Explainable and Robust Deep Learning for Liver Segmentation Through U-Net Network

Background/Objectives: Clinical imaging techniques, such as magnetic resonance imaging and computed tomography, play a vital role in supporting clinicians by aiding disease diagnosis and facilitating the planning of appropriate interventions. This is particularly important in malignant conditions like hepatocellular carcinoma, where accurate image segmentation, delineating the liver and tumor, is a critical initial step in optimizing diagnosis, staging, and treatment planning, including interventions like transplantation, surgical resection, radiotherapy, portal vein embolization, and other procedures. Therefore, effective segmentation methods can significantly influence both diagnostic accuracy and treatment outcomes. Method: In this paper, we propose a deep learning-based approach aimed at accurately segmenting the liver in medical images, thus addressing a critical need in hepatic disease diagnosis and treatment planning. We consider a U-Net architecture with residual connections to capture fine-grained anatomical details. We also take into account the prediction explainability, aiming to highlight, in the image under analysis, the areas that are symptomatic for a certain segmentation. In detail, by exploiting the U-Net architecture, two different models are trained with two annotated datasets of computed tomography medical images, resulting in four different experiments. Results: We consider two different datasets to improve robustness and generalization across diverse patient populations and imaging conditions. Experimental results demonstrate that the proposed method obtains interesting performances, with an accuracy ranging from 0.81 to 0.93. Conclusions: We thus show that the proposed method can provide a reliable and efficient solution for automated liver segmentation, promising significant advancements in clinical workflows and precision medicine.