A Modified U-Net for Semantic Segmentation of Liver and Liver Tumors From CT Scans

Medical image analysis algorithms have been quite popular for automating the segmentation of the liver and liver tumours in recent years. A system like this would also lessen radiologists’ workload and subjective interpretations. Deep learning models are commonly used in this field due to their performance improvement over conventional methods of medical image processing. However, deep learning models’ capacity for class imbalance adaptation is rarely studied, particularly in the instance of segmenting liver tumours given their tiny size. Therefore, in the current study, a straightforward Convolutional Neural Network (CNN) model based on the well-known U-Net architecture is suggested in order to capture more contextual information and solve the class imbalance issue. Additionally, a combination of the loss function is intended to guarantee precise segmentation of liver tumours. A detailed qualitative and quantitative analysis is carried out on LiTs dataset to validate the proposed method.