Implementation and Performance Evaluation of Various Deep Learning-based Segmentation Approaches for Liver Tumor Detection and Diagnosis in Medical Imaging

Liver tumor occurs due to the abnormally grown cell in the liver. It can be of different shapes, size and types. The tumors should be identified timely to avoid unnecessary risks. Only relying on symptom-based identification is not sufficient to diagnose tumors. Different advanced techniques are being applied in medical field to measure the risk factors of the tumors after considering the affected area. After being modernized with AI, healthcare sector got its advanced height of development. Various automatic and intelligent approaches are being proposed for identifying and predicting multiple types of diseases with efficient and reliable performance. Beside disease detection, segmentation techniques are significantly used to identify infected area for further treatments. In this work, an efficient segmentation approach has been designed and proposed for liver tumor diagnosis where six models (Feature Pyramid Network, UNet++, UNet, FCN, Mask R CNN, SegNet) were trained and picked the most efficient one (FPN) for final framework implementation. The models were trained with good numbers of preprocessed images with hyperparameter tuning where 98.75% of accuracy was achieved with FPN (Feature Pyramid Network) model. The proposed framework comes with excellent performance in terms of high accuracy, efficiency and reliability compared to the other existing state-of-the-art techniques.