CT Scan Analysis of Liver Lesion Segmentation with U-Net-ResNeXt Architecture

Image segmentation in the medical field has resulted in detection of various tumors like brain, intestine, skin and liver etc. Segmentation can be performed manually, but the process may be time consuming and the outcomes might be faulty. Hence, automatic liver tumor segmentation helps to detect heterogeneous tumor and non heterogeneous tumor sizes of the liver. The liver tumor is a grievous disease and needs superior treatment in the initial stages to rescue human life. So, in this study an automated Liver Tumor Segmentation architecture with U-net ResNeXt-50 is proposed which uses LiTS17 dataset that contains 130 abdominal CT images. The proposed architecture follows preprocessing steps including augmentation, windowing, conversion format, and normalization. Followed by the U-Net-ResNeXt-50 architecture to segment out the liver and the tumor. At high level U-Net has the encoder-decoder architecture, where the encoder uses the pre-trained network (ResNeXt-50) and decoder segments out the tumor part of the liver. The proposed model achieved performance accuracy of 99.86 and 90.29 F1-score respectively.