U-Net based deep convolutional neural network models for liver segmentation from CT scan images

Liver segmentation is a critical task for diagnosis, treatment and follow-up processes of liver cancer. Computed Tomography (CT) scans are the common medical image modality for the segmentation task. Liver segmentation is considered a very hard task for many reasons. Medical images are limited for researchers. Liver shape is changing based on the patient position during the CT scan process, and varies from patient to another based on the health conditions. Liver and other organs, for example heart, stomach, and pancreas, share similar gray scale range in CT images. Liver treatment using surgery operations is very critical because liver contains significant amount of blood and the position of liver is very close to critical organs like heart, lungs, stomach, and crucial blood veins. Therefore the accuracy of segmentation is critical to define liver and tumors shape and position especially when the treatment surgery conducted using radio frequency heating or cryoablation needles. In the literature, convolutional neural networks (CNN) have achieved very high accuracy on liver segmentation and the U-Net model is considered the state-of-the-art for the medical image segmentation task. Many researchers have developed CNN models based on U-Net and stacked U-Nets with/without bridged connections. However, CNN models need significant number of labeled samples for training and validation which is not commonly available in the case of liver CT images. The process of generating manual annotated masks for the training samples are time consuming and need involvement of expert clinical doctors. Data augmentation has thus been widely used in boosting the sample size for model training. Using rotation with steps of 15o and horizontal and vertical flipping as augmentation techniques, the lack of dataset and training samples issue is solved. The choice of rotation and flipping because in the real life situations, most of the CT scans recorded while the while patient lies on face down or with 45o, 60o,90o on right side according to the location of the tumor. Nonetheless, such process has brought up a new issue for liver segmentation. For example, due to the augmentation operations of rotation and flipping, the trained model detected part of the heart as a liver when it is on the wrong side of the body. The first part of this research conducted an extensive experimental study of U-Net based model in terms of deeper and wider, and variant bridging and skip-connections in order to give recommendation for using U-Net based models. Top-down and bottom-up approaches were used to construct variations of deeper models, whilst two, three, and four stacked U-Nets were applied to construct the wider U-Net models. The variation of the skip connections between two and three U-Nets are the key factors in the study. The proposed model used 2 bridged U-Nets with three extra skip connections between the U-Nets to overcome the flipping issue. A new loss function based on minimizing the distance between the center of mass between the predicted blobs has also enhanced the liver segmentation accuracy. Finally, the deep-supervision concept was integrated with the new loss functions where the total loss was calculated as the sum of weighted loss functions over each weighted deeply supervision. It has achieved a segmentation accuracy of up to 90%. The proposed model of 2 bridged U-Nets with compound skip-connections and specific number of levels, layers, filters, and image size has increased the accuracy of liver segmentation to ~90% whereas the original U-Net and bridged nets have recorded a segmentation accuracy of ~85%. Although applying extra deeply supervised layers and weighted compound of dice coefficient and centroid loss functions solved the flipping issue with ~93%, there is still a room for improving the accuracy by applying some image enhancement as pre-processing stage