CNN-watershed: A watershed transform with predicted markers for corneal endothelium image segmentation
01 Jul 2021-Biomedical Signal Processing and Control (Elsevier)-Vol. 68, pp 102805
TL;DR: In this paper, the watershed algorithm was used for marker-driven segmentation of corneal endothelial cells and an encoder-decoder convolutional neural network trained in a sliding window set up to predict the probability of cell centers (markers) and cell borders.
Abstract: Quantitive information about corneal endothelium cells’ morphometry is vital for assessing cornea pathologies. Nevertheless, in clinical, everyday routine dominates qualitative assessment based on visual inspection of the microscopy images. Although several systems exist for automatic segmentation of corneal endothelial cells, they exhibit certain limitations. The main one is sensitivity to low contrast and uneven illumination, resulting in over-segmentation. Subsequently, image segmentation results often require manual editing of missing or false cell edges. Therefore, this paper further investigates the problem of corneal endothelium cell segmentation. A fully automatic pipeline is proposed that incorporates the watershed algorithm for marker-driven segmentation of corneal endothelial cells and an encoder-decoder convolutional neural network trained in a sliding window set up to predict the probability of cell centers (markers) and cell borders. The predicted markers are used for watershed segmentation of edge probability maps outputted by a neural network. The proposed method's performance on a heterogeneous dataset comprising four publicly available corneal endothelium image datasets is analyzed. The performance of three convolutional neural network models (i.e., U-Net, SegNet, and W-Net) incorporated in the proposed pipeline is examined. The results of the proposed pipeline are analyzed and compared to the state-of-the-art competitor. The obtained results are promising. Regardless of the convolutional neural model incorporated into the proposed pipeline, it notably outperforms the competitor. The proposed method scored 97.72% of cell detection accuracy, compared to 87.38% achieved by the competitor. The advantage of the introduced method is also apparent for cell size, DICE coefficient, and Modified Hausdorff distance.
TL;DR: The aim of the review is to provide information and advice for practitioners to select the appropriate version of watershed for their problem solving, and to forecast future directions of software development for 3D image segmentation by watershed.
Abstract: Watershed is a widely used image segmentation algorithm. Most researchers understand just an idea of this method: a grayscale image is considered as topographic relief, which is flooded from initial basins. However, frequently they are not aware of the options of the algorithm and the peculiarities of its realizations. There are many watershed implementations in software packages and products. Even if these packages are based on the identical algorithm–watershed, by flooding their outcomes, processing speed, and consumed memory, vary greatly. In particular, the difference among various implementations is noticeable for huge volumetric images; for instance, tomographic 3D images, for which low performance and high memory requirements of watershed might be bottlenecks. In our review, we discuss the peculiarities of algorithms with and without waterline generation, the impact of connectivity type and relief quantization level on the result, approaches for parallelization, as well as other method options. We present detailed benchmarking of seven open-source and three commercial software implementations of marker-controlled watershed for semantic or instance segmentation. We compare those software packages for one synthetic and two natural volumetric images. The aim of the review is to provide information and advice for practitioners to select the appropriate version of watershed for their problem solving. In addition, we forecast future directions of software development for 3D image segmentation by watershed.
TL;DR: In this paper , the authors presented a new deep learning-based approach, which automatically detects brain tumors using Magnetic Resonance (MR) images, using convolutional and fully connected layers of a new residual-CNN model trained from scratch.
Abstract: One of the most dangerous diseases in the world is brain tumors. After the brain tumor destroys healthy tissues in the brain, it multiplies abnormally, causing an increase in the internal pressure in the skull. If this condition is not diagnosed early, it can lead to death. Magnetic Resonance Imaging (MRI) is a diagnostic method frequently used in soft tissues with successful results. This study presented a new deep learning-based approach, which automatically detects brain tumors using Magnetic Resonance (MR) images. Convolutional and fully connected layers of a new Residual-CNN (R-CNN) model trained from scratch were used to extract deep features from MR images. The representation power of the deep feature set was increased with the features extracted from all convolutional layers. Among the deep features extracted, the 100 features with the highest distinctiveness were selected with a new multi-level feature selection algorithm named L1NSR. The best performance in the classification stage was obtained by using the SVM algorithm with the Gaussian kernel. The proposed approach was evaluated on two separate data sets composed of 2-class (healthy and tumor) and 4-class (glioma tumor, meningioma tumor, pituitary tumor, and healthy) datasets. Besides, the proposed approach was compared with other state-of-the-art approaches using the respective datasets. The best classification accuracies for 2-class and 4-class datasets were 98.8% and 96.6%, respectively.
TL;DR: In this article , a new deep learning method was proposed to estimate the corneal parameters (endothelial cell density, coefficient of variation, and hexagonality) from specular images.
Abstract: Corneal guttae, which are the abnormal growth of extracellular matrix in the corneal endothelium, are observed in specular images as black droplets that occlude the endothelial cells. To estimate the corneal parameters (endothelial cell density [ECD], coefficient of variation [CV], and hexagonality [HEX]), we propose a new deep learning method that includes a novel attention mechanism (named fNLA), which helps to infer the cell edges in the occluded areas. The approach first derives the cell edges, then infers the well-detected cells, and finally employs a postprocessing method to fix mistakes. This results in a binary segmentation from which the corneal parameters are estimated. We analyzed 1203 images (500 contained guttae) obtained with a Topcon SP-1P microscope. To generate the ground truth, we performed manual segmentation in all images. Several networks were evaluated (UNet, ResUNeXt, DenseUNets, UNet++, etc.) and we found that DenseUNets with fNLA provided the lowest error: a mean absolute error of 23.16 [cells/mm[Formula: see text]] in ECD, 1.28 [%] in CV, and 3.13 [%] in HEX. Compared with Topcon's built-in software, our error was 3-6 times smaller. Overall, our approach handled notably well the cells affected by guttae, detecting cell edges partially occluded by small guttae and discarding large areas covered by extensive guttae.
TL;DR: This paper proposes a hybrid approach to Image Region Extraction that focuses on automated region proposal and segmentation techniques and analyzes popular techniques such as K-Means Clustering and Watershedding and their effectiveness when deployed in a hybrid environment to be applied to a highly variable dataset.
Abstract: With a wide range of applications, image segmentation is a complex and difficult preprocessing step that plays an important role in automatic visual systems, which accuracy impacts, not only on segmentation results, but directly affects the effectiveness of the follow-up tasks. Despite the many advances achieved in the last decades, image segmentation remains a challenging problem, particularly, the segmenting of color images due to the diverse inhomogeneities of color, textures and shapes present in the descriptive features of the images. In trademark graphic images segmentation, beyond these difficulties, we must also take into account the high noise and low resolution, which are often present. Trademark graphic images can also be very heterogeneous with regard to the elements that make them up, which can be overlapping and with varying lighting conditions. Due to the immense variation encountered in corporate logos and trademark graphic images, it is often difficult to select a single method for extracting relevant image regions in a way that produces satisfactory results. Many of the hybrid approaches that integrate the Watershed and K-Means algorithms involve processing very high quality and visually similar images, such as medical images, meaning that either approach can be tweaked to work on images that follow a certain pattern. Trademark images are totally different from each other and are usually fully colored. Our system solves this difficulty given it is a generalized implementation designed to work in most scenarios, through the use of customizable parameters and completely unbiased for an image type. In this paper, we propose a hybrid approach to Image Region Extraction that focuses on automated region proposal and segmentation techniques. In particular, we analyze popular techniques such as K-Means Clustering and Watershedding and their effectiveness when deployed in a hybrid environment to be applied to a highly variable dataset. The proposed system consists of a multi-stage algorithm that takes as input an RGB image and produces multiple outputs, corresponding to the extracted regions. After preprocessing steps, a K-Means function with random initial centroids and a user-defined value for k is executed over the RGB image, generating a gray-scale segmented image, to which a threshold method is applied to generate a binary mask, containing the necessary information to generate a distance map. Then, the Watershed function is performed over the distance map, using the markers defined by the Connected Component Analysis function that labels regions on 8-way pixel connectivity, ensuring that all regions are correctly found. Finally, individual objects are labelled for extraction through a contour method, based on border following. The achieved results show adequate region extraction capabilities when processing graphical images from different datasets, where the system correctly distinguishes the most relevant visual elements of images with minimal tweaking.
TL;DR: Wang et al. as discussed by the authors used an attention-based deep neural network (ADCNN-32s-G) model to segment brain fMRI dataset, which performed well in segmenting mass fMRI datasets.
Abstract: Functional magnetic resonance imaging (fMRI) is widely used for clinical examinations, diagnosis, and treatment. By segmenting fMRI images, large-scale medical image data can be processed more efficiently. Most deep learning (DL)-based segmentation typically uses some type of encoding–decoding model. In this study, affective computing (AC) was developed using the brain fMRI dataset generated from an emotion simulation experiment. The brain fMRI dataset was segmented using an attention model, a deep convolutional neural network-32 (DCNN-32) based on Laplacian of Gaussian (LoG) filter, called ADCNN-32-G. For the evaluation of image segmentation, several indices are presented. By comparing the proposed ADCNN-32s-G model to distance regularized level set evolution (DRLSE), single-seeded region growing, and the single segNet full convolutional network model (FCN), the proposed model performs well in segmenting mass fMRI datasets. The proposed method can be applied to the real-time monitoring of patients with depression, and it can effectively advise human mental health. • Designed brain emotional stimuli experiment. • Used fMRI to explore emotion status of brain. • Developed an attention-based deep neural network to segment brain fMRI dataset. • Decreased computational complexity facing mass medical images dataset in emotion recognition research. • Refers potential applications for the proposed methods in brain research science.
••05 Oct 2015
TL;DR: Neber et al. as discussed by the authors proposed a network and training strategy that relies on the strong use of data augmentation to use the available annotated samples more efficiently, which can be trained end-to-end from very few images and outperforms the prior best method (a sliding-window convolutional network) on the ISBI challenge for segmentation of neuronal structures in electron microscopic stacks.
Abstract: There is large consent that successful training of deep networks requires many thousand annotated training samples. In this paper, we present a network and training strategy that relies on the strong use of data augmentation to use the available annotated samples more efficiently. The architecture consists of a contracting path to capture context and a symmetric expanding path that enables precise localization. We show that such a network can be trained end-to-end from very few images and outperforms the prior best method (a sliding-window convolutional network) on the ISBI challenge for segmentation of neuronal structures in electron microscopic stacks. Using the same network trained on transmitted light microscopy images (phase contrast and DIC) we won the ISBI cell tracking challenge 2015 in these categories by a large margin. Moreover, the network is fast. Segmentation of a 512x512 image takes less than a second on a recent GPU. The full implementation (based on Caffe) and the trained networks are available at http://lmb.informatik.uni-freiburg.de/people/ronneber/u-net .
TL;DR: In this article, the adaptive estimates of lower-order moments are used for first-order gradient-based optimization of stochastic objective functions, based on adaptive estimate of lowerorder moments.
Abstract: We introduce Adam, an algorithm for first-order gradient-based optimization of stochastic objective functions, based on adaptive estimates of lower-order moments. The method is straightforward to implement, is computationally efficient, has little memory requirements, is invariant to diagonal rescaling of the gradients, and is well suited for problems that are large in terms of data and/or parameters. The method is also appropriate for non-stationary objectives and problems with very noisy and/or sparse gradients. The hyper-parameters have intuitive interpretations and typically require little tuning. Some connections to related algorithms, on which Adam was inspired, are discussed. We also analyze the theoretical convergence properties of the algorithm and provide a regret bound on the convergence rate that is comparable to the best known results under the online convex optimization framework. Empirical results demonstrate that Adam works well in practice and compares favorably to other stochastic optimization methods. Finally, we discuss AdaMax, a variant of Adam based on the infinity norm.
TL;DR: The MorphoLibJ library proposes a large collection of generic tools based on MM to process binary and grey-level 2D and 3D images, integrated into user-friendly plugins.
Abstract: Motivation: Mathematical morphology (MM) provides many powerful operators for processing 2D and 3D images. However, most MM plugins currently implemented for the popular ImageJ/Fiji platform are limited to the processing of 2D images. Results: The MorphoLibJ library proposes a large collection of generic tools based on MM to process binary and grey-level 2D and 3D images, integrated into user-friendly plugins. We illustrate how MorphoLibJ can facilitate the exploitation of 3D images of plant tissues. Availability and Implementation: MorphoLibJ is freely available at http://imagej.net/MorphoLibJ
TL;DR: Simulations performed using synthetic generated histograms and a real image show the speed advantage and the accuracy of the iterated version of the derived method for minimum cross entropy thresholding.
Abstract: A fast iterative method is derived for minimum cross entropy thresholding using a one-point iteration scheme. Simulations performed using synthetic generated histograms and a real image show the speed advantage and the accuracy of the iterated version.
TL;DR: SegNet as mentioned in this paper uses pooling indices computed in the max-pooling step of the corresponding encoder to perform non-linear upsampling, which eliminates the need for learning to upsample.
Abstract: We present a novel and practical deep fully convolutional neural network architecture for semantic pixel-wise segmentation termed SegNet. This core trainable segmentation engine consists of an encoder network, a corresponding decoder network followed by a pixel-wise classification layer. The architecture of the encoder network is topologically identical to the 13 convolutional layers in the VGG16 network. The role of the decoder network is to map the low resolution encoder feature maps to full input resolution feature maps for pixel-wise classification. The novelty of SegNet lies is in the manner in which the decoder upsamples its lower resolution input feature map(s). Specifically, the decoder uses pooling indices computed in the max-pooling step of the corresponding encoder to perform non-linear upsampling. This eliminates the need for learning to upsample. The upsampled maps are sparse and are then convolved with trainable filters to produce dense feature maps. We compare our proposed architecture with the widely adopted FCN and also with the well known DeepLab-LargeFOV, DeconvNet architectures. This comparison reveals the memory versus accuracy trade-off involved in achieving good segmentation performance. SegNet was primarily motivated by scene understanding applications. Hence, it is designed to be efficient both in terms of memory and computational time during inference. It is also significantly smaller in the number of trainable parameters than other competing architectures. We also performed a controlled benchmark of SegNet and other architectures on both road scenes and SUN RGB-D indoor scene segmentation tasks. We show that SegNet provides good performance with competitive inference time and more efficient inference memory-wise as compared to other architectures. We also provide a Caffe implementation of SegNet and a web demo at this http URL.