Harnessing Deep Learning and Transfer Learning Models for Segmentation of Liver Tumors and Veins

Liver is one of the most crucial organs in the human body. It helps in bile secretion which aids digestion and metabolism. Healthy liver is important for enhancing immunity of human body. It facilitates the detoxification process and toxin removal from body. Detecting any anomaly in its functioning and structure is important which can be due to diseases and abnormal situations. Tumor is abnormal growth of cells or mass that have chances to be cancerous. But segregating tumors and veins in livers is challenging. It needs expert and experienced eyes to be detected. In automated setup, it becomes more difficult to distinguish between liver veins and tumor without manual intervention. Deep Learning and Computer vision are good techniques to explore their efficacy in the field of Computer Aided Diagnosis for liver disease detection. Segmentation of liver veins and tumors has been a promising task to many researchers. In this paper some deep learning approaches have been used to perform segmentation. Four deep learning models such as Attention Net, Res-U-Net, U-Net and Cascade-Net has been applied to perform segmentation on CT scans of the liver into two different classes, veins in the liver and tumor/s present in the liver. Cascade U-Net achieved highest Accuracy of 95 %, Precision of 0.92, recall of 0.93, specificity of 0.96, dice coefficient of 0.92 and jaccard coefficient of 0.86. Hence the model proved to be a good choice for applicability to medical domain. This paper used some of the light weight models to segment liver tumours and veins with good accuracy, which could be used for real-time implementation as it will be computationally less expensive.