Hybrid Deep Learning Methodology for Automated Hepatic Blood Vessel and Liver Tumor Analysis using CT Imaging

Abstract

Segmentation of hepatic blood vessels and liver tumors, from computed tomography (CT) images is a crucial task in liver cancer diagnosis and surgical planning, vascular assessment, liver transplantation, and monitoring of liver cancer treatment. Although deep learning has seen many developments recently, current methods for liver segmentation fail to concurrently maintain high-resolution hepatic vascular structures while accurately segmenting tumors with variable shapes, sizes and contrast. While Conventional Convolutional Neural Network (CNN) models focus mainly on learning local spatial representations, Transformer models have a stronger focus on learning global contextual information, resulting in a lack of ability to comprehensively capture the complementary characteristics of features within a single framework. Considering the above challenges, this study proposes a novel Hybrid Deep Learning Methodology for Automated Hepatic Blood Vessel and Liver Tumor Analysis Using CT Imaging (HDLM-HTA). The proposed framework leverages the benefits of multi-scale convolutional feature extraction, dual attention-based feature refinement, global context learning with Vision Transformer, adaptive feature fusion, and multi-task segmentation in a single end-to-end pipeline to improve segmentation accuracy and robustness.The pre-processing stage of the proposed methodology consists of several steps: intensity normalization, bilateral filtering, extraction of Region of Interest (ROI) of the liver, resizing the CT image to the standard resolution and data augmentation to increase the image quality and promote the generalization of the model. Various receptive fields of a multi-scale CNN can then extract hierarchical spatial features, and a dual attention mechanism (channel attention and spatial attention) can highlight clinically relevant vessel and tumor regions to enhance the discriminative anatomical representation. Then, a global context learning module based on Vision Transformer is implemented to make use of the long-range anatomical dependency and guarantee vascular continuity. Using an adaptive feature fusion strategy, local features learnt from the affected liver tissue are fused with the globally learnt features and sent to the multi-task segmentation decoder, which is responsible for generating the hepatic blood vessel and liver tumour masks simultaneously. Optimization of the network is done by minimizing a hybrid loss function that consists of Dice Loss, Binary Cross-Entropy Loss and Focal Loss, which is suitable for class imbalance, yet still enhances segmentation performance.Experimental tests demonstrate that the proposed HDLM-HTA framework greatly improves the current CNN and hybrid deep learning models on various quantitative metrics including Dice Similarity Coefficient, Intersection over Union, Precision, Recall, F1-score, Hausdorff Distance and Average Surface Distance. It is a significant improvement to obtain multi-scale feature extraction, dual attention mechanism, transformer-based contextual learning, and adaptive feature fusion, which are essential for retaining fine branches of liver vessels and for defining the borders of complex liver tumors. These results show that the proposed automated liver image analysis framework is reliable, robust and clinically applicable. In summary, the proposed approach of HDLM-HTA is an effective computer-aided liver blood vessels segmentation and liver tumor segmentation method that can be used for liver cancer diagnosis, surgical planning, vascular mapping, treatment monitoring and precision medicine applications.