Multi-Class Segmentation of Liver & Tumor Using 2D U-Net with ResNet152 as Backbone

One of the main causes of deaths among adults throughout the world is liver cancer. It is difficult and exhaustively time-consuming to identify liver lesion from healthy tissues and cancer tissue from non-cancerous lesion manually in the present scenario of diagnosis which makes medical image segmentation to be very crucial to separate out the liver and corresponding tumor from the CT images of patients' abdomen to diagnose tumor. Numerous studies were done on the topic. With the advancement of technology, autonomous segmentation of liver and tumor has been possible from image datasets of CT or MRI. Though researchers have made significant improvement in this field, there is still scope of enhancing the accuracy of liver and tumor segmentation. The aim of the study is to perform segmentation on liver and tumor classes simultaneously from CT images and enhance the accuracy of segmentation of both liver and tumor. Data augmentation using elastic deformation was performed on the training dataset along with few other augmentation techniques. 2D U-Net model incorporating ResNet152 as the backbone was trained to obtain the segmentation of liver and liver tumor. Dice scores of 96.53% and 81±O.09% are achieved for liver and liver tumor segmentation respectively. The volume overlap error for both liver and tumor are significantly reduced which indicates to a remarkable improvement for the study. The study will help in the clinical practices and decision making in future.