The Adaptive Attention Fusion Network for Automated Liver and Tumor Segmentation in CT Images

Abstract

In the modern treatment of hepatic diseases, liver parenchyma and hepatic malignancies have to be accurately segmented from computed tomography (CT) scans for such applications as surgical resection planning or radiation therapy. Medical image segmentation has made great progress; however automated extraction is still limited due to the high degree of volumetric class imbalance between the background tissues and the liver organ, as well as small tumors and between tissues and organs with unclear boundaries and morphology. One of the major technical challenges in the current deep learning methods is that they simply apply a one-size-fits-all solution by directly combining noisy shallow encoder features with deep decoder semantics, which causes a lack of boundary degradation. To close this basic void, the Adaptive Attention Fusion Network, which incorporates a novel Adaptive Attention Fusion Module, has been proposed in this research work. The ultimate goal is to move away from this discrete approach of skipping and introduce an approach that gets a signal at each spatial location, dynamically assessing its contextual relevance for each point, and accentuating the clinical signal while downscaling the background noise. A dataset for the empirical evaluation was created using the 3D-IRCADb-01, containing 2,823 axial slices for 20 patient volumes, all adequately divided at the patient level from the previous ones, to avoid data leakage. The methodological pipeline consisted of liver-specific Hounsfield Unit windowing, stochastic geometric augmentations, and application of a composite loss function with a weighted cross-entropy and multi-class Dice loss. The network was then rigorously evaluated against the standard U-Net, Attention U-Net, and DeepLabV3+ models with a fixed test split and a 3-fold cross-validation framework across four different metrics, viz., Dice, Intersection over Union, Precision, and Recall. The findings show that the proposed network could attain state-of-the-art segmentation performance with regard to the liver, by obtaining a high score of 0.8126 with Dice, 0.7589 with Intersection over Union, and an extremely high score of 0.9597 in the sense of Recall, which demonstrates strong capability to remove any kind of ambiguity in the boundaries and eliminate any unwanted false-negative omission. The Dice score of segmentation of the single split was 0.7048, and the expanded 3-fold cross-validation gave an increased mean tumor Dice score of 0.7639, leading to better long-term geometric generalization. The mathematical representation of the continuous pixel-wise gating mechanism to define an improved paradigm for spatial feature integration is the main contribution. To summarize, this framework manages to accommodate the theoretical computational design aspects with the necessary clinical need and utility, offering an efficient screening and assistive computational diagnostic tool in a contemporary context of oncological imaging.