Unbox the black-box for the medical explainable AI via multi-modal and multi-centre data fusion: A mini-review, two showcases and beyond

Convolutional neural networks (CNNs) have revolutionized medical image analysis over the past few years. The U-Net architecture is one of the most well-known CNN architectures for semantic segmentation and has achieved remarkable successes in many different medical image segmentation applications. The U-Net architecture consists of standard convolution layers, pooling layers, and upsampling layers. These convolution layers learn representative features of input images and construct segmentations based on the features. However, the features learned by standard convolution layers are not distinctive when the differences among different categories are subtle in terms of intensity, location, shape, and size. In this paper, we propose a novel CNN architecture, called Dense-Res-Inception Net (DRINet), which addresses this challenging problem. The proposed DRINet consists of three blocks, namely a convolutional block with dense connections, a deconvolutional block with residual inception modules, and an unpooling block. Our proposed architecture outperforms the U-Net in three different challenging applications, namely multi-class segmentation of cerebrospinal fluid on brain CT images, multi-organ segmentation on abdominal CT images, and multi-class brain tumor segmentation on MR images.

/pdf/drinet-for-medical-image-segmentation-4ld69ovke7.pdf

DRINet for Medical Image Segmentation

Distance regularized two level sets for segmentation of left and right ventricles from cine-MRI.

Automated Segmentation of Tissues Using CT and MRI: A Systematic Review.

Stroke is a serious manifestation of various cerebrovascular diseases and one of the most dangerous diseases in the world today. Volume quantification and location detection of chronic stroke lesions provide vital biomarkers for stroke rehabilitation. Recently, deep learning has seen a rapid growth, with a great potential in segmenting medical images. In this work, unlike most deep learning-based segmentation methods utilizing only magnetic resonance (MR) images as the input, we propose and validate a novel stroke lesion segmentation approach named multi-inputs UNet (MI-UNet) that incorporates brain parcellation information, including gray matter (GM), white matter (WM) and lateral ventricle (LV). The brain parcellation is obtained from 3D diffeomorphic registration and is concatenated with the original MR image to form two-channel inputs to the subsequent MI-UNet. Effectiveness of the proposed pipeline is validated using a dataset consisting of 229 T1-weighted MR images. Experiments are conducted via a five-fold cross-validation. The proposed MI-UNet performed significantly better than UNet in both 2D and 3D settings. Our best results obtained by 3D MI-UNet has superior segmentation performance, as measured by the Dice score, Hausdorff distance, average symmetric surface distance, as well as precision, over other state-of-the-art methods.

MI-UNet: Multi-Inputs UNet Incorporating Brain Parcellation for Stroke Lesion Segmentation From T1-Weighted Magnetic Resonance Images

Purpose: Myocardium segmentation in cardiac magnetic resonance (MR) images plays a vital role in clinical diagnosis of the cardiovascular diseases. Because of the low contrast and large variation in intensity and shapes, myocardium segmentation has been a challenging task. A dynamic programming (DP) based segmentation method, incorporating the likelihood and shape information of the myocardium, is developed for segmenting myocardium in cardiac MR images. Methods: The endocardium, i.e., the left ventricle blood cavity, is segmented for initialization, and then the optimal epicardium contour is determined using the polar-transformed image and DP scheme. In the DP segmentation scheme, three techniques are proposed to improve the segmentation performance: (1) the likelihood image of the myocardium is constructed to define the external cost in the DP, thus the cost function incorporates prior probability estimation; (2) the adaptive search range is introduced to determine the polar-transformed image, thereby excluding irrelevant tissues; (3) the connectivity constrained DP algorithm is developed to obtain an optimal closed contour. Four metrics, including the Dice metric (Dice), root mean squared error (RMSE), reliability, and correlation coefficient, are used to assess the segmentation accuracy. The authors evaluated the performance of the proposed method on a private dataset and the MICCAI 2009 challenge dataset. The authors also explored the effects of the three new techniques of the DP scheme in the proposed method. Results: For the qualitative evaluation, the segmentation results of the proposed method were clinically acceptable. For the quantitative evaluation, the mean (Dice) for the endocardium and epicardium was 0.892 and 0.927, respectively; the mean RMSE was 2.30 mm for the endocardium and 2.39 mm for the epicardium. In addition, the three new techniques in the proposed DP scheme, i.e., the likelihood image of the myocardium, the adaptive search range, and the connectivity constrained DP algorithm, improved the segmentation performance for the epicardium with 0.029, 0.047, and 0.007 in terms of the Dice and 0.98, 1.31, and 0.21 mm in terms of the RMSE, respectively. Conclusions: The three techniques (the likelihood image of the myocardium, the adaptive search range, and the connectivity constrained DP algorithm) can improve the segmentation ability of the DP method, and the proposed method with these techniques has the ability to achieve the acceptable segmentation result of the myocardium in cardiac MR images. Therefore, the proposed method would be useful in clinical diagnosis of the cardiovascular diseases.

Segmentation of myocardium from cardiac MR images using a novel dynamic programming based segmentation method.

Ventricle segmentation is a challenging technique for the development of detection system of ischemic stroke in computed tomography (CT), as ischemic stroke regions are adjacent to the brain ventricle with similar intensity. To address this problem, we developed an objective segmentation system of brain ventricle in CT. The intensity distribution of the ventricle was estimated based on clustering technique, connectivity, and domain knowledge, and the initial ventricle segmentation results were then obtained. To exclude the stroke regions from initial segmentation, a combined segmentation strategy was proposed, which is composed of three different schemes: (1) the largest three-dimensional (3D) connected component was considered as the ventricular region; (2) the big stroke areas were removed by the image difference methods based on searching optimal threshold values; (3) the small stroke regions were excluded by the adaptive template algorithm. The proposed method was evaluated on 50 cases of patients with ischemic stroke. The mean Dice, sensitivity, specificity, and root mean squared error were 0.9447, 0.969, 0.998, and 0.219 mm, respectively. This system can offer a desirable performance. Therefore, the proposed system is expected to bring insights into clinic research and the development of detection system of ischemic stroke in CT.

/pdf/objective-ventricle-segmentation-in-brain-ct-with-ischemic-3s90wmqy3h.pdf

Yue Zhao

Papers

Segmentation of myocardium from cardiac MR images using a novel dynamic programming based segmentation method.

Objective Ventricle Segmentation in Brain CT with Ischemic Stroke Based on Anatomical Knowledge.