Explainable AI- Driven Hybrid Framework for Liver Fibrosis Segmentation

Abstract

Accurate segmentation of cirrhotic liver tissue in T2-weighted MRI is important for quantitative analysis, diagnosis aid and treatment planning. However, this task is challenging for conventional segmentation models due to the presence of weak boundaries, heterogeneous texture and shape variability. Motivated by these difficulties, the present thesis proposes a hybrid deep learning framework which combines transformer-based global context modelling with convolutional feature extraction to improve segmentation performance and reliability. The proposed approach combines SegFormer-B2 and ResNet-50 in a single encoderdecoder architecture for 2D liver segmentation. Hyperparameter optimization was done using Optuna. The final model was trained with a combined loss function including binary cross-entropy, Dice loss, and boundary-aware supervision. To enhance interpretability, the model also generated Grad-CAM based explainability maps to visualize the regions influencing model predictions. The model evaluated the model on the CirrMRI600+ T2W 2D benchmark and compared it to several strong baselines. The experimental results indicated that the proposed model obtained a Dice score of 92.67% and mIoU of 86.57% on the test set, and the best validation Dice was 93.33%. The results presented here show that the hybrid architecture is able to capture semantic context as well as fine anatomical detail, resulting in accurate and robust liver segmentation. In addition, the explainability analysis confirmed that the model focused on clinically relevant regions, which increased trust in its predictions. Overall, the research shows that hybrid transformer-convolution architectures, combined with boundary-aware training and explainable AI techniques, offer a promising solution for reliable cirrhotic liver MRI segmentation.