Showing papers in "Medical Physics in 2017"

A deep convolutional neural network using directional wavelets for low-dose X-ray CT reconstruction.

Abstract: Purpose Due to the potential risk of inducing cancer, radiation exposure by X-ray CT devices should be reduced for routine patient scanning However, in low-dose X-ray CT, severe artifacts typically occur due to photon starvation, beam hardening, and other causes, all of which decrease the reliability of the diagnosis Thus, a high-quality reconstruction method from low-dose X-ray CT data has become a major research topic in the CT community Conventional model-based de-noising approaches are, however, computationally very expensive, and image-domain de-noising approaches cannot readily remove CT-specific noise patterns To tackle these problems, we want to develop a new low-dose X-ray CT algorithm based on a deep-learning approach Method We propose an algorithm which uses a deep convolutional neural network (CNN) which is applied to the wavelet transform coefficients of low-dose CT images More specifically, using a directional wavelet transform to extract the directional component of artifacts and exploit the intra- and inter- band correlations, our deep network can effectively suppress CT-specific noise In addition, our CNN is designed with a residual learning architecture for faster network training and better performance Results Experimental results confirm that the proposed algorithm effectively removes complex noise patterns from CT images derived from a reduced X-ray dose In addition, we show that the wavelet-domain CNN is efficient when used to remove noise from low-dose CT compared to existing approaches Our results were rigorously evaluated by several radiologists at the Mayo Clinic and won second place at the 2016 “Low-Dose CT Grand Challenge” Conclusions To the best of our knowledge, this work is the first deep-learning architecture for low-dose CT reconstruction which has been rigorously evaluated and proven to be effective In addition, the proposed algorithm, in contrast to existing model-based iterative reconstruction (MBIR) methods, has considerable potential to benefit from large data sets Therefore, we believe that the proposed algorithm opens a new direction in the area of low-dose CT research

MR-based synthetic CT generation using a deep convolutional neural network method

Abstract: Purpose Interests have been rapidly growing in the field of radiotherapy to replace CT with magnetic resonance imaging (MRI), due to superior soft tissue contrast offered by MRI and the desire to reduce unnecessary radiation dose. MR-only radiotherapy also simplifies clinical workflow and avoids uncertainties in aligning MR with CT. Methods, however, are needed to derive CT-equivalent representations, often known as synthetic CT (sCT), from patient MR images for dose calculation and DRR-based patient positioning. Synthetic CT estimation is also important for PET attenuation correction in hybrid PET-MR systems. We propose in this work a novel deep convolutional neural network (DCNN) method for sCT generation and evaluate its performance on a set of brain tumor patient images. Methods The proposed method builds upon recent developments of deep learning and convolutional neural networks in the computer vision literature. The proposed DCNN model has 27 convolutional layers interleaved with pooling and unpooling layers and 35 million free parameters, which can be trained to learn a direct end-to-end mapping from MR images to their corresponding CTs. Training such a large model on our limited data is made possible through the principle of transfer learning and by initializing model weights from a pretrained model. Eighteen brain tumor patients with both CT and T1-weighted MR images are used as experimental data and a sixfold cross-validation study is performed. Each sCT generated is compared against the real CT image of the same patient on a voxel-by-voxel basis. Comparison is also made with respect to an atlas-based approach that involves deformable atlas registration and patch-based atlas fusion. Results The proposed DCNN method produced a mean absolute error (MAE) below 85 HU for 13 of the 18 test subjects. The overall average MAE was 84.8 ± 17.3 HU for all subjects, which was found to be significantly better than the average MAE of 94.5 ± 17.8 HU for the atlas-based method. The DCNN method also provided significantly better accuracy when being evaluated using two other metrics: the mean squared error (188.6 ± 33.7 versus 198.3 ± 33.0) and the Pearson correlation coefficient(0.906 ± 0.03 versus 0.896 ± 0.03). Although training a DCNN model can be slow, training only need be done once. Applying a trained model to generate a complete sCT volume for each new patient MR image only took 9 s, which was much faster than the atlas-based approach. Conclusions A DCNN model method was developed, and shown to be able to produce highly accurate sCT estimations from conventional, single-sequence MR images in near real time. Quantitative results also showed that the proposed method competed favorably with an atlas-based method, in terms of both accuracy and computation speed at test time. Further validation on dose computation accuracy and on a larger patient cohort is warranted. Extensions of the method are also possible to further improve accuracy or to handle multi-sequence MR images.

Use of image registration and fusion algorithms and techniques in radiotherapy: Report of the AAPM Radiation Therapy Committee Task Group No. 132.

Abstract: Image registration and fusion algorithms exist in almost every software system that creates or uses images in radiotherapy. Most treatment planning systems support some form of image registration and fusion to allow the use of multimodality and time-series image data and even anatomical atlases to assist in target volume and normal tissue delineation. Treatment delivery systems perform registration and fusion between the planning images and the in-room images acquired during the treatment to assist patient positioning. Advanced applications are beginning to support daily dose assessment and enable adaptive radiotherapy using image registration and fusion to propagate contours and accumulate dose between image data taken over the course of therapy to provide up-to-date estimates of anatomical changes and delivered dose. This information aids in the detection of anatomical and functional changes that might elicit changes in the treatment plan or prescription. As the output of the image registration process is always used as the input of another process for planning or delivery, it is important to understand and communicate the uncertainty associated with the software in general and the result of a specific registration. Unfortunately, there is no standard mathematical formalism to perform this for real-world situations where noise, distortion, and complex anatomical variations can occur. Validation of the software systems performance is also complicated by the lack of documentation available from commercial systems leading to use of these systems in undesirable ‘black-box’ fashion. In view of this situation and the central role that image registration and fusion play in treatment planning and delivery, the Therapy Physics Committee of the American Association of Physicists in Medicine commissioned Task Group 132 to review current approaches and solutions for image registration (both rigid and deformable) in radiotherapy and to provide recommendations for quality assurance and quality control of these clinical processes.

Intrinsic dependencies of CT radiomic features on voxel size and number of gray levels.

Abstract: Purpose Many radiomics features were originally developed for non-medical imaging applications and therefore original assumptions may need to be reexamined. In this study, we investigated the impact of slice thickness and pixel spacing (or pixel size) on radiomics features extracted from Computed Tomography (CT) phantom images acquired with different scanners as well as different acquisition and reconstruction parameters. The dependence of CT texture features on gray-level discretization was also evaluated. Methods and materials A texture phantom composed of 10 different cartridges of different materials was scanned on eight different CT scanners from three different manufacturers. The images were reconstructed for various slice thicknesses. For each slice thickness, the reconstruction Field Of View (FOV) was varied to render pixel sizes ranging from 0.39 to 0.98 mm. A fixed spherical region of interest (ROI) was contoured on the images of the shredded rubber cartridge and the 3D printed, 20% fill, acrylonitrile butadiene styrene plastic cartridge (ABS20) for all phantom imaging sets. Radiomic features were extracted from the ROIs using an in-house program. Features categories were: shape (10), intensity (16), GLCM (24), GLZSM (11), GLRLM (11), and NGTDM (5), fractal dimensions (8) and first-order wavelets (128), for a total of 213 features. Voxel-size resampling was performed to investigate the usefulness of extracting features using a suitably chosen voxel size. Acquired phantom image sets were resampled to a voxel size of 1 × 1 × 2 mm3 using linear interpolation. Image features were therefore extracted from resampled and original datasets and the absolute value of the percent coefficient of variation (%COV) for each feature was calculated. Based on the %COV values, features were classified in 3 groups: (1) features with large variations before and after resampling (%COV >50); (2) features with diminished variation (%COV <30) after resampling; and (3) features that had originally moderate variation (%COV <50%) and were negligibly affected by resampling. Group 2 features were further studied by modifying feature definitions to include voxel size. Original and voxel-size normalized features were used for interscanner comparisons. A subsequent analysis investigated feature dependency on gray-level discretization by extracting 51 texture features from ROIs from each of the 10 different phantom cartridges using 16, 32, 64, 128, and 256 gray levels. Results Out of the 213 features extracted, 150 were reproducible across voxel sizes, 42 improved significantly (%COV <30, Group 2) after resampling, and 21 had large variations before and after resampling (Group 1). Ten features improved significantly after definition modification effectively removed their voxel-size dependency. Interscanner comparison indicated that feature variability among scanners nearly vanished for 8 of these 10 features. Furthermore, 17 out of 51 texture features were found to be dependent on the number of gray levels. These features were redefined to include the number of gray levels which greatly reduced this dependency. Conclusion Voxel-size resampling is an appropriate pre-processing step for image datasets acquired with variable voxel sizes to obtain more reproducible CT features. We found that some of the radiomics features were voxel size and gray-level discretization-dependent. The introduction of normalizing factors in their definitions greatly reduced or removed these dependencies.

Segmentation of organs-at-risks in head and neck CT images using convolutional neural networks

Abstract: Purpose Accurate segmentation of organs-at-risks (OARs) is the key step for efficient planning of radiation therapy for head and neck (HaN) cancer treatment. In the work, we proposed the first deep learning-based algorithm, for segmentation of OARs in HaN CT images, and compared its performance against state-of-the-art automated segmentation algorithms, commercial software, and interobserver variability. Methods Convolutional neural networks (CNNs)—a concept from the field of deep learning—were used to study consistent intensity patterns of OARs from training CT images and to segment the OAR in a previously unseen test CT image. For CNN training, we extracted a representative number of positive intensity patches around voxels that belong to the OAR of interest in training CT images, and negative intensity patches around voxels that belong to the surrounding structures. These patches then passed through a sequence of CNN layers that captured local image features such as corners, end-points, and edges, and combined them into more complex high-order features that can efficiently describe the OAR. The trained network was applied to classify voxels in a region of interest in the test image where the corresponding OAR is expected to be located. We then smoothed the obtained classification results by using Markov random fields algorithm. We finally extracted the largest connected component of the smoothed voxels classified as the OAR by CNN, performed dilate–erode operations to remove cavities of the component, which resulted in segmentation of the OAR in the test image. Results The performance of CNNs was validated on segmentation of spinal cord, mandible, parotid glands, submandibular glands, larynx, pharynx, eye globes, optic nerves, and optic chiasm using 50 CT images. The obtained segmentation results varied from 37.4% Dice coefficient (DSC) for chiasm to 89.5% DSC for mandible. We also analyzed the performance of state-of-the-art algorithms and commercial software reported in the literature, and observed that CNNs demonstrate similar or superior performance on segmentation of spinal cord, mandible, parotid glands, larynx, pharynx, eye globes, and optic nerves, but inferior performance on segmentation of submandibular glands and optic chiasm. Conclusion We concluded that convolution neural networks can accurately segment most of OARs using a representative database of 50 HaN CT images. At the same time, inclusion of additional information, for example, MR images, may be beneficial to some OARs with poorly visible boundaries.

A deep feature fusion methodology for breast cancer diagnosis demonstrated on three imaging modality datasets.

Abstract: Background Deep learning methods for radiomics/computer-aided diagnosis (CADx) are often prohibited by small datasets, long computation time, and the need for extensive image preprocessing. Aims We aim to develop a breast CADx methodology that addresses the aforementioned issues by exploiting the efficiency of pre-trained convolutional neural networks (CNNs) and using pre-existing handcrafted CADx features. Materials & Methods We present a methodology that extracts and pools low- to mid-level features using a pretrained CNN and fuses them with handcrafted radiomic features computed using conventional CADx methods. Our methodology is tested on three different clinical imaging modalities (dynamic contrast enhanced-MRI [690 cases], full-field digital mammography [245 cases], and ultrasound [1125 cases]). Results From ROC analysis, our fusion-based method demonstrates, on all three imaging modalities, statistically significant improvements in terms of AUC as compared to previous breast cancer CADx methods in the task of distinguishing between malignant and benign lesions. (DCE-MRI [AUC = 0.89 (se = 0.01)], FFDM [AUC = 0.86 (se = 0.01)], and ultrasound [AUC = 0.90 (se = 0.01)]). Discussion/Conclusion We proposed a novel breast CADx methodology that can be used to more effectively characterize breast lesions in comparison to existing methods. Furthermore, our proposed methodology is computationally efficient and circumvents the need for image preprocessing.

Automatic segmentation of the clinical target volume and organs at risk in the planning CT for rectal cancer using deep dilated convolutional neural networks.

Abstract: Purpose Delineation of the clinical target volume (CTV) and organs at risk (OARs) is very important for radiotherapy but is time-consuming and prone to inter-observer variation Here, we proposed a novel deep dilated convolutional neural network (DDCNN)-based method for fast and consistent auto-segmentation of these structures Methods Our DDCNN method was an end-to-end architecture enabling fast training and testing Specifically, it employed a novel multiple-scale convolutional architecture to extract multiple-scale context features in the early layers, which contain the original information on fine texture and boundaries and which are very useful for accurate auto-segmentation In addition, it enlarged the receptive fields of dilated convolutions at the end of networks to capture complementary context features Then, it replaced the fully connected layers with fully convolutional layers to achieve pixel-wise segmentation We used data from 278 patients with rectal cancer for evaluation The CTV and OARs were delineated and validated by senior radiation oncologists in the planning computed tomography (CT) images A total of 218 patients chosen randomly were used for training, and the remaining 60 for validation The Dice similarity coefficient (DSC) was used to measure segmentation accuracy Results Performance was evaluated on segmentation of the CTV and OARs In addition, the performance of DDCNN was compared with that of U-Net The proposed DDCNN method outperformed the U-Net for all segmentations, and the average DSC value of DDCNN was 38% higher than that of U-Net Mean DSC values of DDCNN were 877% for the CTV, 934% for the bladder, 921% for the left femoral head, 923% for the right femoral head, 653% for the intestine and 618% for the colon The test time was 45 s per patient for segmentation of all the CTV, bladder, left and right femoral heads, colon and intestine We also assessed our approaches and results with those in the literature: our system showed superior performance and faster speed Conclusions These data suggest that DDCNN can be used to segment the CTV and OARs accurately and efficiently It was invariant to the body size, body shape, and age of the patients DDCNN could improve the consistency of contouring and streamline radiotherapy workflows This article is protected by copyright All rights reserved

Evaluation of segmentation methods on head and neck CT: Auto-segmentation challenge 2015

Abstract: Purpose Automated delineation of structures and organs is a key step in medical imaging. However, due to the large number and diversity of structures and the large variety of segmentation algorithms, a consensus is lacking as to which automated segmentation method works best for certain applications. Segmentation challenges are a good approach for unbiased evaluation and comparison of segmentation algorithms. Methods In this work, we describe and present the results of the Head and Neck Auto-Segmentation Challenge 2015, a satellite event at the Medical Image Computing and Computer Assisted Interventions (MICCAI) 2015 conference. Six teams participated in a challenge to segment nine structures in the head and neck region of CT images: brainstem, mandible, chiasm, bilateral optic nerves, bilateral parotid glands, and bilateral submandibular glands. Results This paper presents the quantitative results of this challenge using multiple established error metrics and a well-defined ranking system. The strengths and weaknesses of the different auto-segmentation approaches are analyzed and discussed. Conclusions The Head and Neck Auto-Segmentation Challenge 2015 was a good opportunity to assess the current state-of-the-art in segmentation of organs at risk for radiotherapy treatment. Participating teams had the possibility to compare their approaches to other methods under unbiased and standardized circumstances. The results demonstrate a clear tendency toward more general purpose and fewer structure-specific segmentation algorithms.

AAPM TG 158: Measurement and calculation of doses outside the treated volume from external-beam radiation therapy.

Abstract: The introduction of advanced techniques and technology in radiotherapy has greatly improved our ability to deliver highly conformal tumor doses while minimizing the dose to adjacent organs at risk. Despite these tremendous improvements, there remains a general concern about doses to normal tissues that are not the target of the radiation treatment; any “nontarget” radiation should be minimized as it offers no therapeutic benefit. As patients live longer after treatment, there is increased opportunity for late effects including second cancers and cardiac toxicity to manifest. Complicating the management of these issues, there are unique challenges with measuring, calculating, reducing, and reporting nontarget doses that many medical physicists may have limited experience with. Treatment planning systems become dramatically inaccurate outside the treatment field, necessitating a measurement or some other means of assessing the dose. However, measurements are challenging because outside the treatment field, the radiation energy spectrum, dose rate, and general shape of the dose distribution (particularly the percent depth dose) are very different and often require special consideration. Neutron dosimetry is also particularly challenging, and common errors in methodology can easily manifest as errors of several orders of magnitude. Task Group 158 was, therefore, formed to provide guidance for physicists in terms of assessing and managing nontarget doses. In particular, the report: (a) highlights major concerns with nontarget radiation; (b) provides a rough estimate of doses associated with different treatment approaches in clinical practice; (c) discusses the uses of dosimeters for measuring photon, electron, and neutron doses; (d) discusses the use of calculation techniques for dosimetric evaluations; (e) highlights techniques that may be considered for reducing nontarget doses; (f) discusses dose reporting; and (g) makes recommendations for both clinical and research practice.

Using deep learning to segment breast and fibroglandular tissue in MRI volumes

Abstract: Purpose Automated segmentation of breast and fibroglandular tissue (FGT) is required for various computer-aided applications of breast MRI. Traditional image analysis and computer vision techniques, such atlas, template matching, or, edge and surface detection, have been applied to solve this task. However, applicability of these methods is usually limited by the characteristics of the images used in the study datasets, while breast MRI varies with respect to the different MRI protocols used, in addition to the variability in breast shapes. All this variability, in addition to various MRI artifacts, makes it a challenging task to develop a robust breast and FGT segmentation method using traditional approaches. Therefore, in this study, we investigated the use of a deep-learning approach known as "U-net." Materials and methods We used a dataset of 66 breast MRI's randomly selected from our scientific archive, which includes five different MRI acquisition protocols and breasts from four breast density categories in a balanced distribution. To prepare reference segmentations, we manually segmented breast and FGT for all images using an in-house developed workstation. We experimented with the application of U-net in two different ways for breast and FGT segmentation. In the first method, following the same pipeline used in traditional approaches, we trained two consecutive (2C) U-nets: first for segmenting the breast in the whole MRI volume and the second for segmenting FGT inside the segmented breast. In the second method, we used a single 3-class (3C) U-net, which performs both tasks simultaneously by segmenting the volume into three regions: nonbreast, fat inside the breast, and FGT inside the breast. For comparison, we applied two existing and published methods to our dataset: an atlas-based method and a sheetness-based method. We used Dice Similarity Coefficient (DSC) to measure the performances of the automated methods, with respect to the manual segmentations. Additionally, we computed Pearson's correlation between the breast density values computed based on manual and automated segmentations. Results The average DSC values for breast segmentation were 0.933, 0.944, 0.863, and 0.848 obtained from 3C U-net, 2C U-nets, atlas-based method, and sheetness-based method, respectively. The average DSC values for FGT segmentation obtained from 3C U-net, 2C U-nets, and atlas-based methods were 0.850, 0.811, and 0.671, respectively. The correlation between breast density values based on 3C U-net and manual segmentations was 0.974. This value was significantly higher than 0.957 as obtained from 2C U-nets (P < 0.0001, Steiger's Z-test with Bonferoni correction) and 0.938 as obtained from atlas-based method (P = 0.0016). Conclusions In conclusion, we applied a deep-learning method, U-net, for segmenting breast and FGT in MRI in a dataset that includes a variety of MRI protocols and breast densities. Our results showed that U-net-based methods significantly outperformed the existing algorithms and resulted in significantly more accurate breast density computation.

Deep reinforcement learning for automated radiation adaptation in lung cancer.

Abstract: Purpose To investigate deep reinforcement learning (DRL) based on historical treatment plans for developing automated radiation adaptation protocols for nonsmall cell lung cancer (NSCLC) patients that aim to maximize tumor local control at reduced rates of radiation pneumonitis grade 2 (RP2). Methods In a retrospective population of 114 NSCLC patients who received radiotherapy, a three-component neural networks framework was developed for deep reinforcement learning (DRL) of dose fractionation adaptation. Large-scale patient characteristics included clinical, genetic, and imaging radiomics features in addition to tumor and lung dosimetric variables. First, a generative adversarial network (GAN) was employed to learn patient population characteristics necessary for DRL training from a relatively limited sample size. Second, a radiotherapy artificial environment (RAE) was reconstructed by a deep neural network (DNN) utilizing both original and synthetic data (by GAN) to estimate the transition probabilities for adaptation of personalized radiotherapy patients' treatment courses. Third, a deep Q-network (DQN) was applied to the RAE for choosing the optimal dose in a response-adapted treatment setting. This multicomponent reinforcement learning approach was benchmarked against real clinical decisions that were applied in an adaptive dose escalation clinical protocol. In which, 34 patients were treated based on avid PET signal in the tumor and constrained by a 17.2% normal tissue complication probability (NTCP) limit for RP2. The uncomplicated cure probability (P+) was used as a baseline reward function in the DRL. Results Taking our adaptive dose escalation protocol as a blueprint for the proposed DRL (GAN + RAE + DQN) architecture, we obtained an automated dose adaptation estimate for use at ∼2/3 of the way into the radiotherapy treatment course. By letting the DQN component freely control the estimated adaptive dose per fraction (ranging from 1-5 Gy), the DRL automatically favored dose escalation/de-escalation between 1.5 and 3.8 Gy, a range similar to that used in the clinical protocol. The same DQN yielded two patterns of dose escalation for the 34 test patients, but with different reward variants. First, using the baseline P+ reward function, individual adaptive fraction doses of the DQN had similar tendencies to the clinical data with an RMSE = 0.76 Gy; but adaptations suggested by the DQN were generally lower in magnitude (less aggressive). Second, by adjusting the P+ reward function with higher emphasis on mitigating local failure, better matching of doses between the DQN and the clinical protocol was achieved with an RMSE = 0.5 Gy. Moreover, the decisions selected by the DQN seemed to have better concordance with patients eventual outcomes. In comparison, the traditional temporal difference (TD) algorithm for reinforcement learning yielded an RMSE = 3.3 Gy due to numerical instabilities and lack of sufficient learning. Conclusion We demonstrated that automated dose adaptation by DRL is a feasible and a promising approach for achieving similar results to those chosen by clinicians. The process may require customization of the reward function if individual cases were to be considered. However, development of this framework into a fully credible autonomous system for clinical decision support would require further validation on larger multi-institutional datasets.

Classification and evaluation strategies of auto‐segmentation approaches for PET: Report of AAPM task group No. 211

Abstract: Purpose The purpose of this educational report is to provide an overview of the present state-of-the-art PET auto-segmentation (PET-AS) algorithms and their respective validation, with an emphasis on providing the user with help in understanding the challenges and pitfalls associated with selecting and implementing a PET-AS algorithm for a particular application. Approach A brief description of the different types of PET-AS algorithms is provided using a classification based on method complexity and type. The advantages and the limitations of the current PET-AS algorithms are highlighted based on current publications and existing comparison studies. A review of the available image datasets and contour evaluation metrics in terms of their applicability for establishing a standardized evaluation of PET-AS algorithms is provided. The performance requirements for the algorithms and their dependence on the application, the radiotracer used and the evaluation criteria are described and discussed. Finally, a procedure for algorithm acceptance and implementation, as well as the complementary role of manual and auto-segmentation are addressed. Findings A large number of PET-AS algorithms have been developed within the last 20 years. Many of the proposed algorithms are based on either fixed or adaptively selected thresholds. More recently, numerous papers have proposed the use of more advanced image analysis paradigms to perform semi-automated delineation of the PET images. However, the level of algorithm validation is variable and for most published algorithms is either insufficient or inconsistent which prevents recommending a single algorithm. This is compounded by the fact that realistic image configurations with low signal-to-noise ratios (SNR) and heterogeneous tracer distributions have rarely been used. Large variations in the evaluation methods used in the literature point to the need for a standardized evaluation protocol. Conclusions Available comparison studies suggest that PET-AS algorithms relying on advanced image analysis paradigms provide generally more accurate segmentation than approaches based on PET activity thresholds, particularly for realistic configurations. However, this may not be the case for simple shape lesions in situations with a narrower range of parameters, where simpler methods may also perform well. Recent algorithms which employ some type of consensus or automatic selection between several PET-AS methods have potential to overcome the limitations of the individual methods when appropriately trained. In either case, accuracy evaluation is required for each different PET scanner and scanning and image reconstruction protocol. For the simpler, less robust approaches, adaptation to scanning conditions, tumor type, and tumor location by optimization of parameters is necessary. The results from the method evaluation stage can be used to estimate the contouring uncertainty. All PET-AS contours should be critically verified by a physician. A standard test, i.e., a benchmark dedicated to evaluating both existing and future PET-AS algorithms needs to be designed, to aid clinicians in evaluating and selecting PET-AS algorithms and to establish performance limits for their acceptance for clinical use. The initial steps toward designing and building such a standard are undertaken by the task group members.

Development of the open-source dose calculation and optimization toolkit matRad.

Abstract: Purpose We report on the development of the open-source cross-platform radiation treatment planning toolkit matRad and its comparison against validated treatment planning systems. The toolkit enables three-dimensional intensity-modulated radiation therapy treatment planning for photons, scanned protons and scanned carbon ions. Methods matRad is entirely written in Matlab and is freely available online. It re-implements well-established algorithms employing a modular and sequential software design to model the entire treatment planning workflow. It comprises core functionalities to import DICOM data, to calculate and optimize dose as well as a graphical user interface for visualization. matRad dose calculation algorithms (for carbon ions this also includes the computation of the relative biological effect) are compared against dose calculation results originating from clinically approved treatment planning systems. Results We observe three-dimensional γ-analysis pass rates ≥ 99.67% for all three radiation modalities utilizing a distance to agreement of 2 mm and a dose difference criterion of 2%. The computational efficiency of matRad is evaluated in a treatment planning study considering three different treatment scenarios for every radiation modality. For photons, we measure total run times of 145 s–1260 s for dose calculation and fluence optimization combined considering 4–72 beam orientations and 2608–13597 beamlets. For charged particles, we measure total run times of 63 s–993 s for dose calculation and fluence optimization combined considering 9963–45574 pencil beams. Using a CT and dose grid resolution of 0.3 cm3 requires a memory consumption of 1.59 GB–9.07 GB and 0.29 GB–17.94 GB for photons and charged particles, respectively. Conclusion The dosimetric accuracy, computational performance and open-source character of matRad encourages a future application of matRad for both educational and research purposes.

A parallel MR imaging method using multilayer perceptron.

Abstract: Purpose To reconstruct MR images from subsampled data, we propose a fast reconstruction method using the multilayer perceptron (MLP) algorithm Methods and Materials We applied MLP to reduce aliasing artifacts generated by subsampling in k-space The MLP is learned from training data to map aliased input images into desired alias-free images The input of the MLP is all voxels in the aliased lines of multi-channel real and imaginary images from the subsampled k-space data, and the desired output is all voxels in the corresponding alias-free line of the root-sum-of-squares of multi-channel images from fully-sampled k-space data Aliasing artifacts in an image reconstructed from subsampled data were reduced by line-by-line processing of the learned MLP architecture Results Reconstructed images from the proposed method are better than those from compared methods in terms of normalized root-mean-square error The proposed method can be applied to image reconstruction for any k-space subsampling patterns in a phase encoding direction Moreover, to further reduce the reconstruction time, it is easily implemented by parallel processing Conclusion We have proposed a reconstruction method using machine learning to accelerate imaging time, which reconstructs high-quality images from subsampled k-space data It shows flexibility in the use of k-space sampling patterns, and can reconstruct images in real time This article is protected by copyright All rights reserved

Deep learning for polyp recognition in wireless capsule endoscopy images.

Abstract: Purpose Wireless capsule endoscopy (WCE) enables physicians to examine the digestive tract without any surgical operations, at the cost of a large volume of images to be analyzed. In the computer-aided diagnosis of WCE images, the main challenge arises from the difficulty of robust characterization of images. This study aims to provide discriminative description of WCE images and assist physicians to recognize polyp images automatically. Methods We propose a novel deep feature learning method, named stacked sparse autoencoder with image manifold constraint (SSAEIM), to recognize polyps in the WCE images. Our SSAEIM differs from the traditional sparse autoencoder (SAE) by introducing an image manifold constraint, which is constructed by a nearest neighbor graph and represents intrinsic structures of images. The image manifold constraint enforces that images within the same category share similar learned features and images in different categories should be kept far away. Thus, the learned features preserve large intervariances and small intravariances among images. Results The average overall recognition accuracy (ORA) of our method for WCE images is 98.00%. The accuracies for polyps, bubbles, turbid images, and clear images are 98.00%, 99.50%, 99.00%, and 95.50%, respectively. Moreover, the comparison results show that our SSAEIM outperforms existing polyp recognition methods with relative higher ORA. Conclusion The comprehensive results have demonstrated that the proposed SSAEIM can provide descriptive characterization for WCE images and recognize polyps in a WCE video accurately. This method could be further utilized in the clinical trials to help physicians from the tedious image reading work.

Low-dose CT for the detection and classification of metastatic liver lesions: Results of the 2016 Low Dose CT Grand Challenge.

Abstract: Purpose Using common datasets, to estimate and compare the diagnostic performance of image-based denoising techniques or iterative reconstruction algorithms for the task of detecting hepatic metastases. Methods Datasets from contrast-enhanced CT scans of the liver were provided to participants in an NIH-, AAPM- and Mayo Clinic-sponsored Low Dose CT Grand Challenge. Training data included full-dose and quarter-dose scans of the ACR CT accreditation phantom and 10 patient examinations; both images and projections were provided in the training data. Projection data were supplied in a vendor-neutral standardized format (DICOM-CT-PD). Twenty quarter-dose patient datasets were provided to each participant for testing the performance of their technique. Images were provided to sites intending to perform denoising in the image domain. Fully preprocessed projection data and statistical noise maps were provided to sites intending to perform iterative reconstruction. Upon return of the denoised or iteratively reconstructed quarter-dose images, randomized, blinded evaluation of the cases was performed using a Latin Square study design by 11 senior radiology residents or fellows, who marked the locations of identified hepatic metastases. Markings were scored against reference locations of clinically or pathologically demonstrated metastases to determine a per-lesion normalized score and a per-case normalized score (a faculty abdominal radiologist established the reference location using clinical and pathological information). Scores increased for correct detections; scores decreased for missed or incorrect detections. The winner for the competition was the entry that produced the highest total score (mean of the per-lesion and per-case normalized score). Reader confidence was used to compute a Jackknife alternative free-response receiver operating characteristic (JAFROC) figure of merit, which was used for breaking ties. Results 103 participants from 90 sites and 26 countries registered to participate. Training data were shared with 77 sites that completed the data sharing agreements. Subsequently, 41 sites downloaded the 20 test cases, which included only the 25% dose data (CTDIvol = 3.0 ± 1.8 mGy, SSDE = 3.5 ± 1.3 mGy). 22 sites submitted results for evaluation. One site provided binary images and one site provided images with severe artifacts; cases from these sites were excluded from review and the participants removed from the challenge. The mean (range) per-lesion and per-case normalized scores were −24.2% (−75.8%, 3%) and 47% (10%, 70%), respectively. Compared to reader results for commercially reconstructed quarter-dose images with no noise reduction, 11 of the 20 sites showed a numeric improvement in the mean JAFROC figure of merit. Notably two sites performed comparably to the reader results for full-dose commercial images. The study was not designed for these comparisons, so wide confidence intervals surrounded these figures of merit and the results should be used only to motivate future testing. Conclusion Infrastructure and methodology were developed to rapidly estimate observer performance for liver metastasis detection in low-dose CT examinations of the liver after either image-based denoising or iterative reconstruction. The results demonstrated large differences in detection and classification performance between noise reduction methods, although the majority of methods provided some improvement in performance relative to the commercial quarter-dose images with no noise reduction applied.

Deep learning of the sectional appearances of 3D CT images for anatomical structure segmentation based on an FCN voting method.

Abstract: Purpose We propose a single network trained by pixel-to-label deep learning to address the general issue of automatic multiple organ segmentation in three-dimensional (3D) computed tomography (CT) images. Our method can be described as a voxel-wise multiple-class classification scheme for automatically assigning labels to each pixel/voxel in a 2D/3D CT image. Methods We simplify the segmentation algorithms of anatomical structures (including multiple organs) in a CT image (generally in 3D) to a majority voting scheme over the semantic segmentation of multiple 2D slices drawn from different viewpoints with redundancy. The proposed method inherits the spirit of fully convolutional networks (FCNs) that consist of “convolution” and “deconvolution” layers for 2D semantic image segmentation, and expands the core structure with 3D-2D-3D transformations to adapt to 3D CT image segmentation. All parameters in the proposed network are trained pixel-to-label from a small number of CT cases with human annotations as the ground truth. The proposed network naturally fulfills the requirements of multiple organ segmentations in CT cases of different sizes that cover arbitrary scan regions without any adjustment. Results The proposed network was trained and validated using the simultaneous segmentation of 19 anatomical structures in the human torso, including 17 major organs and two special regions (lumen and content inside of stomach). Some of these structures have never been reported in previous research on CT segmentation. A database consisting of 240 (95% for training and 5% for testing) 3D CT scans, together with their manually annotated ground-truth segmentations, was used in our experiments. The results show that the 19 structures of interest were segmented with acceptable accuracy (88.1% and 87.9% voxels in the training and testing datasets, respectively, were labeled correctly) against the ground truth. Conclusions We propose a single network based on pixel-to-label deep learning to address the challenging issue of anatomical structure segmentation in 3D CT cases. The novelty of this work is the policy of deep learning of the different 2D sectional appearances of 3D anatomical structures for CT cases and the majority voting of the 3D segmentation results from multiple crossed 2D sections to achieve availability and reliability with better efficiency, generality, and flexibility than conventional segmentation methods, which must be guided by human expertise. This article is protected by copyright. All rights reserved.

Photon-counting CT for simultaneous imaging of multiple contrast agents in the abdomen: An in vivo study

Abstract: Purpose To demonstrate the feasibility of spectral imaging using photon-counting detector (PCD) x-ray computed tomography (CT) for simultaneous material decomposition of three contrast agents in vivo in a large animal model. Methods This Institutional Animal Care and Use Committee-approved study used a canine model. Bismuth subsalicylate was administered orally 24–72 h before imaging. PCD CT was performed during intravenous administration of 40–60 ml gadoterate meglumine; 3.5 min later, iopamidol 370 was injected intravenously. Renal PCD CT images were acquired every 2 s for 5–6 min to capture the wash-in and wash-out kinetics of the contrast agents. Least mean squares linear material decomposition was used to calculate the concentrations of contrast agents in the aorta, renal cortex, renal medulla and renal pelvis. Results Using reference vials with known concentrations of materials, we computed molar concentrations of the various contrast agents during each phase of CT scanning. Material concentration maps allowed simultaneous quantification of both arterial and delayed renal enhancement in a single CT acquisition. The accuracy of the material decomposition algorithm in a test phantom was −0.4 ± 2.2 mM, 0.3 ± 2.2 mM for iodine and gadolinium solutions, respectively. Peak contrast concentration of gadolinium and iodine in the aorta, renal cortex, and renal medulla were observed 16, 24, and 60 s after the start each injection, respectively. Conclusion Photon-counting spectral CT allowed simultaneous material decomposition of multiple contrast agents in vivo. Besides defining contrast agent concentrations, tissue enhancement at multiple phases was observed in a single CT acquisition, potentially obviating the need for multiphase CT scans and thus reducing radiation dose.

High dose-per-pulse electron beam dosimetry - A model to correct for the ion recombination in the Advanced Markus ionization chamber.

Abstract: The purpose of this work was to establish an empirical model of the ion recombination in the Advanced Markus ionization chamber for measurements in high dose rate/dose-per-pulse electron beams. In addition, we compared the observed ion recombination to calculations using the standard Boag two-voltage-analysis method, the more general theoretical Boag models, and the semiempirical general equation presented by Burns and McEwen. Two independent methods were used to investigate the ion recombination: (a) Varying the grid tension of the linear accelerator (linac) gun (controls the linac output) and measuring the relative effect the grid tension has on the chamber response at different source-to-surface distances (SSD). (b) Performing simultaneous dose measurements and comparing the dose-response, in beams with varying dose rate/dose-per-pulse, with the chamber together with dose rate/dose-per-pulse independent Gafchromic™ EBT3 film. Three individual Advanced Markus chambers were used for the measurements with both methods. All measurements were performed in electron beams with varying mean dose rate, dose rate within pulse, and dose-per-pulse (10(-2) ≤ mean dose rate ≤ 10(3) Gy/s, 10(2) ≤ mean dose rate within pulse ≤ 10(7) Gy/s, 10(-4) ≤ dose-per-pulse ≤ 10(1) Gy), which was achieved by independently varying the linac gun grid tension, and the SSD. The results demonstrate how the ion collection efficiency of the chamber decreased as the dose-per-pulse increased, and that the ion recombination was dependent on the dose-per-pulse rather than the dose rate, a behavior predicted by Boag theory. The general theoretical Boag models agreed well with the data over the entire investigated dose-per-pulse range, but only for a low polarizing chamber voltage (50 V). However, the two-voltage-analysis method and the Burns & McEwen equation only agreed with the data at low dose-per-pulse values (≤ 10(-2) and ≤ 10(-1) Gy, respectively). An empirical model of the ion recombination in the chamber was found by fitting a logistic function to the data. The ion collection efficiency of the Advanced Markus ionization chamber decreases for measurements in electron beams with increasingly higher dose-per-pulse. However, this chamber is still functional for dose measurements in beams with dose-per-pulse values up toward and above 10 Gy, if the ion recombination is taken into account. Our results show that existing models give a less-than-accurate description of the observed ion recombination. This motivates the use of the presented empirical model for measurements with the Advanced Markus chamber in high dose-per-pulse electron beams, as it enables accurate absorbed dose measurements (uncertainty estimation: 2.8-4.0%, k = 1). The model depends on the dose-per-pulse in the beam, and it is also influenced by the polarizing chamber voltage, with increasing ion recombination with a lowering of the voltage.

Motion management strategies and technical issues associated with stereotactic body radiotherapy of thoracic and upper abdominal tumors: A review from NRG oncology

Abstract: The efficacy of stereotactic body radiotherapy (SBRT) has been well demonstrated. However, it presents unique challenges for accurate planning and delivery especially in the lungs and upper abdomen where respiratory motion can be significantly confounding accurate targeting and avoidance of normal tissues. In this paper, we review the current literature on SBRT for lung and upper abdominal tumors with particular emphasis on addressing respiratory motion and its affects. We provide recommendations on strategies to manage motion for different, patient-specific situations. Some of the recommendations will potentially be adopted to guide clinical trial protocols.

The potential of dual-energy CT to reduce proton beam range uncertainties.

Abstract: PURPOSE: Dual‐energy CT (DECT) promises improvements in estimating stopping power ratios (SPRs) for proton therapy treatment planning. Although several comparable mathematical formalisms have been proposed in literature, the optimal techniques to characterize human tissue SPRs with DECT in a clinical environment are not fully established. The aim of this work is to compare the most robust DECT methods against conventional single‐energy CT (SECT) in conditions reproducing a clinical environment, where CT artifacts and noise play a major role on the accuracy of these techniques.METHODS: Available DECT tissue characterization methods are investigated and their ability to predict SPRs is compared in three contexts: (a) a theoretical environment using the XCOM cross section database; (b) experimental data using a dual‐source CT scanner on a calibration phantom; (c) simulations of a virtual humanoid phantom with the ImaSim software. The latter comparison accounts for uncertainties caused by CT artifacts and noise, but leaves aside other sources of uncertainties such as CT grid size and the I‐values. To evaluate the clinical impact, a beam range calculation model is used to predict errors from the probability distribution functions determined with ImaSim simulations. Range errors caused by SPR errors in soft tissues and bones are investigated. RESULTS: Range error estimations demonstrate that DECT has the potential of reducing proton beam range uncertainties by 0.4% in soft tissues using low noise levels of 12 and 8 HU in DECT, corresponding to 7 HU in SECT. For range uncertainties caused by the transport of protons through bones, the reduction in range uncertainties for the same levels of noise is found to be up to 0.6 to 1.1 mm for bone thicknesses ranging from 1 to 5 cm, respectively. We also show that for double the amount noise, i.e., 14 HU in SECT and 24 and 16 HU for DECT, the advantages of DECT in soft tissues are lost over SECT. In bones however, the reduction in range uncertainties is found to be between 0.5 and 0.9 mm for bone thicknesses ranging from 1 to 5 cm, respectively. CONCLUSION: DECT has a clear potential to improve proton beam range predictions over SECT in proton therapy. However, in the current state high levels of noise remain problematic for DECT characterization methods and do not allow getting the full benefits of this technology. Future work should focus on adapting DECT methods to noise and investigate methods based on raw‐data to reduce CT artifacts.

High dose-per-pulse electron beam dosimetry: Usability and dose-rate independence of EBT3 Gafchromic films.

Abstract: Purpose The aim of this study was to assess the suitability of Gafchromic EBT3 films for reference dose measurements in the beam of a prototype high dose-per-pulse linear accelerator (linac), capable of delivering electron beams with a mean dose-rate (Ḋm) ranging from 0.07 to 3000 Gy/s and a dose-rate in pulse (Ḋp) of up to 8·106 Gy/s. To do this, we evaluated the overall uncertainties in EBT3 film dosimetry as well as the energy and dose-rate dependence of their response. Material and Methods Our dosimetric system was composed of EBT3 Gafchromic films in combination with a flatbed scanner and was calibrated against an ionization chamber traceable to primary standard. All sources of uncertainties in EBT3 dosimetry were carefully analyzed using irradiations at a clinical radiotherapy linac. Energy dependence was investigated with the same machine by acquiring and comparing calibration curves for three different beam energies (4, 8 and 12 MeV), for doses between 0.25 and 30 Gy. Ḋm dependence was studied at the clinical linac by changing the pulse repetition frequency (f) of the beam in order to vary Ḋm between 0.55 and 4.40 Gy/min, while Ḋp dependence was probed at the prototype machine for Ḋp ranging from 7·103 to 8·106 Gy/s. Ḋp dependence was first determined by studying the correlation between the dose measured by films and the charge of electrons measured at the exit of the machine by an induction torus. Furthermore, we compared doses from the films to independently calibrated thermo-luminescent dosimeters (TLD) that have been reported as being dose-rate independent up to such high dose-rates. Results We report that uncertainty below 4% (k=2) can be achieved in the dose range between 3 and 17 Gy. Results also demonstrated that EBT3 films did not display any detectable energy dependence for electron beam energies between 4 and 12 MeV. No Ḋm dependence was found either. In addition, we obtained excellent consistency between films and TLDs over the entire Ḋp range attainable at the prototype linac confirming the absence of any dose-rate dependence within the investigated range (7·103 to 8·106 Gy/s). This aspect was further corroborated by the linear relationship between the dose per pulse (Dp) measured by films and the charge per pulse (Cp) measured at the prototype linac exit. Conclusion Our study shows that the use of EBT3 Gafchromic films can be extended to reference dosimetry in pulsed electron beams with a very high dose-rate. The measurement results are associated with an overall uncertainty below 4% (k=2) and are dose-rate and energy independent. This article is protected by copyright. All rights reserved.

Cascade convolutional neural networks for automatic detection of thyroid nodules in ultrasound images

Abstract: Purpose It is very important for calculation of clinical indices and diagnosis to detect thyroid nodules from ultrasound images. However, this task is a challenge mainly due to heterogeneous thyroid nodules with distinct components are similar to background in ultrasound images. In this study, we employ cascade deep convolutional neural networks (CNNs) to develop and evaluate a fully automatic detection of thyroid nodules from 2D ultrasound images. Methods Our cascade CNNs are a type of hybrid model, consisting of two different CNNs and a new splitting method. Specifically, it employs a deep CNN to learn the segmentation probability maps from the ground true data. Then, all the segmentation probability maps are split into different connected regions by the splitting method. Finally, another deep CNN is used to automatically detect the thyroid nodules from ultrasound thyroid images. Results Experiment results illustrate the cascade CNNs are very effective in detection of thyroid nodules. Specially, the value of area under the curve of receiver operating characteristic is 98.51%. The Free-response receiver operating characteristic (FROC) and jackknife alternative FROC (JAFROC) analyses show a significant improvement in the performance of our cascade CNNs compared to that of other methods. The multi-view strategy can improve the performance of cascade CNNs. Moreover, our special splitting method can effectively separate different connected regions so that the second CNN can correctively gain the positive and negative samples according to the automatic labels. Conclusions The experiment results demonstrate the potential clinical applications of this proposed method. This technique can offer physicians an objective second opinion, and reduce their heavy workload so as to avoid misdiagnosis causes because of excessive fatigue. In addition, it is easy and reproducible for a person without medical expertise to diagnose thyroid nodules.

Future of medical physics: Real-time MRI-guided proton therapy

Abstract: With the recent clinical implementation of real-time MRI-guided x-ray beam therapy (MRXT), attention is turning to the concept of combining real-time MRI guidance with proton beam therapy; MRI-guided proton beam therapy (MRPT). MRI guidance for proton beam therapy is expected to offer a compelling improvement to the current treatment workflow which is warranted arguably more than for x-ray beam therapy. This argument is born out of the fact that proton therapy toxicity outcomes are similar to that of the most advanced IMRT treatments, despite being a fundamentally superior particle for cancer treatment. In this Future of Medical Physics article, we describe the various software and hardware aspects of potential MRPT systems and the corresponding treatment workflow. Significant software developments, particularly focused around adaptive MRI-based planning will be required. The magnetic interaction between the MRI and the proton beamline components will be a key area of focus. For example, the modeling and potential redesign of a magnetically compatible gantry to allow for beam delivery from multiple angles towards a patient located within the bore of an MRI scanner. Further to this, the accuracy of pencil beam scanning and beam monitoring in the presence of an MRI fringe field will require modeling, testing, and potential further development to ensure that the highly targeted radiotherapy is maintained. Looking forward we envisage a clear and accelerated path for hardware development, leveraging from lessons learnt from MRXT development. Within few years, simple prototype systems will likely exist, and in a decade, we could envisage coupled systems with integrated gantries. Such milestones will be key in the development of a more efficient, more accurate, and more successful form of proton beam therapy for many common cancer sites.

Dual-energy CT quantitative imaging: a comparison study between twin-beam and dual-source CT scanners.

Abstract: Purpose: To assess image quality and to quantify the accuracy of relative electron densities (rho(e)) and effective atomic numbers (Z(eff)) for three dual-energy computed tomography (DECT) scanners: a novel single-source split-filter (i. e., twin-beam) and two dual-source scanners. Methods: Measurements were made with a second generation dual-source scanner at 80/140Sn kVp, a third-generation twin-beam single-source scanner at 120 kVp with gold (Au) and tin (Sn) filters, and a third-generation dual-source scanner at 90/150Sn kVp. Three phantoms with tissue inserts were scanned and used for calibration and validation of parameterized methods to extract rho(e) and Z(eff), whereas iodine and calcium inserts were used to quantify Contrast-to-Noise-Ratio (CNR). Spatial resolution in tomographic images was also tested. Results: The third-generation scanners have an image resolution of 6.2, similar to 0.5 lp/cm higher than the second generation scanner. The twin-beam scanner has low imaging contrast for iodine materials due to its limited spectral separation. The parameterization methods resulted in calibrations with low fit residuals for the dual-source scanners, yielding values of rho(e) and Z(eff) close to the reference values (errors within 1.2% for rho(e) and 6.2% for Z(eff) for a dose of 20 mGy, excluding lung substitute tissues). The twin-beam scanner presented overall higher errors (within 3.2% for rho(e) and 28% for Z(eff), also excluding lung inserts) and also larger variations for uniform inserts. Conclusions: Spatial resolution is similar for the three scanners. The twin-beam is able to derive qe and Z(eff), but with inferior accuracy compared to both dual-source scanners. (C) 2016 American Association of Physicists in Medicine

Discriminating solitary cysts from soft tissue lesions in mammography using a pretrained deep convolutional neural network.

Abstract: Purpose: It is estimated that 7% of women in the western world will develop palpable breast cysts in their lifetime. Even though cysts have been correlated with risk of developing breast cancer, many of them are benign and do not require follow-up. We develop a method to discriminate benign solitary cysts from malignant masses in digital mammography. We think a system like this can have merit in the clinic as a decision aid or complementary to specialized modalities. Methods: We employ a deep convolutional neural network (CNN) to classify cyst and mass patches. Deep CNNs have been shown to be powerful classifiers, but need a large amount of training data for which medical problems are often difficult to come by. The key contribution of this paper is that we show good performance can be obtained on a small dataset by pretraining the network on a large dataset of a related task. We subsequently investigate the following: (a) when a mammographic exam is performed, two different views of the same breast are recorded. We investigate the merit of combining the output of the classifier from these two views. (b) We evaluate the importance of the resolution of the patches fed to the network. (c) A method dubbed tissue augmentation is subsequently employed, where we extract normal tissue from normal patches and superimpose this onto the actual samples aiming for a classifier invariant to occluding tissue. (d) We combine the representation extracted using the deep CNN with our previously developed features. Results: We show that using the proposed deep learning method, an area under the ROC curve (AUC) value of 0.80 can be obtained on a set of benign solitary cysts and malignant mass findings recalled in screening. We find that it works significantly better than our previously developed approach by comparing the AUC of the ROC using bootstrapping. By combining views, the results can be further improved, though this difference was not found to be significant. We find no significant difference between using a resolution of 100 versus 200 micron. The proposed tissue augmentations give a small improvement in performance, but this improvement was also not found to be significant. The final system obtained an AUC of 0.80 with 95% confidence interval [0.78, 0.83], calculated using bootstrapping. The system works best for lesions larger than 27 mm where it obtains an AUC value of 0.87. Conclusion: We have presented a computer-aided diagnosis (CADx) method to discriminate cysts from solid lesion in mammography using features from a deep CNN trained on a large set of mass candidates, obtaining an AUC of 0.80 on a set of diagnostic exams recalled from screening. We believe the system shows great potential and comes close to the performance of recently developed spectral mammography. We think the system can be further improved when more data and computational power becomes available. c 2017 American Association of Physicists in Medicine

A computer-aided diagnostic system for detecting diabetic retinopathy in optical coherence tomography images.

Abstract: Purpose Detection (diagnosis) of diabetic retinopathy (DR) in optical coherence tomography (OCT) images for patients with type 2 diabetes, but almost clinically normal retina appearances. Methods The proposed computer-aided diagnostic (CAD) system detects the DR in three steps: (i) localizing and segmenting 12 distinct retinal layers on the OCT image; (ii) deriving features of the segmented layers, and (iii) learning most discriminative features and classifying each subject as normal or diabetic. To localise and segment the retinal layers, signals (intensities) of the OCT image are described with a joint Markov-Gibbs random field (MGRF) model of intensities and shape descriptors. Each segmented layer is characterized with cumulative probability distribution functions (CDF) of its locally extracted features, such as reflectivity, curvature, and thickness. A multistage deep fusion classification network (DFCN) with a stack of non-negativity-constrained autoencoders (NCAE) is trained to select the most discriminative retinal layers’ features and use their CDFs for detecting the DR. A training atlas was built using the OCT scans for 12 normal subjects and their maps of layers hand-drawn by retina experts. Results Preliminary experiments on 52 clinical OCT scans (26 normal and 26 with early-stage DR, balanced between 40{79 years old males and females; 40 training and 12 test subjects) gave the DR detection accuracy, sensitivity, and specificity of 92%; 83%, and 100%, respectively. The 100% accuracy, sensitivity, and specificity have been obtained in the leave-one-out cross-validation test for all the 52 subjects. Conclusion Both the quantitative and visual assessments confirmed the high accuracy of the proposed computer-assisted diagnostic system for early DR detection using the OCT retinal images. This article is protected by copyright. All rights reserved.

Applications of nonlocal means algorithm in low-dose X-ray CT image processing and reconstruction: A review

Abstract: Low-dose X-ray computed tomography (LDCT) imaging is highly recommended for use in the clinic because of growing concerns over excessive radiation exposure. However, the CT images reconstructed by the conventional filtered back-projection (FBP) method from low-dose acquisitions may be severely degraded with noise and streak artifacts due to excessive X-ray quantum noise, or with view-aliasing artifacts due to insufficient angular sampling. In 2005, the nonlocal means (NLM) algorithm was introduced as a non-iterative edge-preserving filter to denoise natural images corrupted by additive Gaussian noise, and showed superior performance. It has since been adapted and applied to many other image types and various inverse problems. This paper specifically reviews the applications of the NLM algorithm in LDCT image processing and reconstruction, and explicitly demonstrates its improving effects on the reconstructed CT image quality from low-dose acquisitions. The effectiveness of these applications on LDCT and their relative performance are described in detail.

REtroSpective Evaluation of Cerebral Tumors (RESECT): a clinical database of pre-operative MRI and intra-operative ultrasound in low-grade glioma surgeries

Abstract: Purpose The advancement of medical image processing techniques, such as image registration, can effectively help improve the accuracy and efficiency of brain tumor surgeries. However, it is often challenging to validate these techniques with real clinical data due to the rarity of such publicly available repositories. Acquisition and validation methods Pre-operative magnetic resonance images (MRI), and intra-operative ultrasound (US) scans were acquired from 23 patients with low-grade glioma who underwent surgeries at St. Olavs University Hospital between 2011 and 2016. Each patient was scanned by Gadolinium-enhanced T1w and T2-FLAIR MRI protocols to reveal the anatomy and pathology, and series of B-mode ultrasound images were obtained before, during, and after tumor resection to track the surgical progress and tissue deformation. Retrospectively, corresponding anatomical landmarks were identified across US images of different surgical stages, and between MRI and US, and can be used to validate image registration algorithms. Quality of landmark identification was assessed with intra- and inter-rater variability. Data format and access In addition to co-registered MRIs, each series of US scans are provided as a reconstructed 3D volume. All images are accessible in MINC2 and NIFTI formats, and the anatomical landmarks were annotated in MNI tag files. Both the imaging data and the corresponding landmarks are available online as the RESECT database at https://archive.norstore.no. Potential impact The proposed database provides real high-quality multi-modal clinical data to validate and compare image registration algorithms that can potentially benefit the accuracy and efficiency of brain tumor resection. Furthermore, the database can also be used to test other image processing methods and neuro-navigation software platforms. This article is protected by copyright. All rights reserved.

Urinary bladder cancer staging in CT urography using machine learning.

Abstract: Purpose To evaluate the feasibility of using an objective computer aided system to assess bladder cancer stage in CT Urography (CTU). Materials and Methods A data set consisting of 84 bladder cancer lesions from 76 CTU cases was used to develop the computerized system for bladder cancer staging based on machine learning approaches. The cases were grouped into two classes based on pathological stage ≥T2 or below T2, which is the decision threshold for neoadjuvant chemotherapy treatment clinically. There were 43 cancers below stage T2 and 41 cancers at stage T2 or above. All 84 lesions were automatically segmented using our previously developed auto-initialized cascaded level sets (AI-CALS) method. Morphological and texture features were extracted. The features were divided into subspaces of morphological features only, texture features only, and a combined set of both morphological and texture features. The data set was split into Set 1 and Set 2 for two-fold cross validation. Stepwise feature selection was used to select the most effective features. A linear discriminant analysis (LDA), a neural network (NN), a support vector machine (SVM), and a random forest (RAF) classifier were used to combine the features into a single score. The classification accuracy of the four classifiers was compared using the area under the receiver operating characteristic (ROC) curve (Az). Results Based on the texture features only, the LDA classifier achieved a test Az of 0.91 on Set 1 and a test Az of 0.88 on Set 2. The test Az of the NN classifier for Set 1 and Set 2 were 0.89 and 0.92, respectively. The SVM classifier achieved test Az of 0.91 on Set 1 and test Az of 0.89 on Set 2. The test Az of the RAF classifier for Set 1 and Set 2 was 0.89 and 0.97, respectively. The morphological features alone, the texture features alone, and the combined feature set achieved comparable classification performance. Conclusion The predictive model developed in this study shows promise as a classification tool for stratifying bladder cancer into two staging categories: greater than or equal to stage T2 and below stage T2. This article is protected by copyright. All rights reserved.