T-MPEDNet: Unveiling the Synergy of Transformer-aware Multiscale Progressive Encoder-Decoder Network with Feature Recalibration for Tumor and Liver Segmentation

Abstract

Early diagnosis and optimal treatment planning are crucial for patients with liver diseases and malignancies. Automated segmentation of the liver and tumors in Computed Tomography (CT) scans and Magnetic Resonance Imaging (MRI) is crucial in surgical planning with improved diagnostic accuracy. However, existing methods face challenges due to the heterogeneous nature of tumors with varying visual characteristics of livers across diverse patient populations. To overcome these challenges, we propose the Transformer-aware Multiscale Progressive Encoder-Decoder Network (T-MPEDNet), which advances automated liver and tumor segmentation. T-MPEDNet integrates a progressive encoder–decoder structure with adaptive skip connections that recalibrate channel-wise features while preserving spatial integrity. A transformer-inspired dynamic attention mechanism captures long-range contextual relationships, complemented by a multi-scale feature extraction module for local detail refinement at varying scales. A morphological boundary refinement module also sharpens indistinct boundaries, ensuring precise liver and tumor segmentation. We demonstrate T-MPEDNet's superiority over seventeen state-of-the-art methods through extensive quantitative and qualitative analyses. We evaluate T-MPEDNet's efficacy on three public benchmark datasets, LiTS, 3DIRCADb, and ATLAS. T-MPEDNet achieves Dice Similarity Coefficients (DSC) of 97.6% (liver) and 89.1% (tumor) on LiTS, 98.3% (liver) and 83.3% (tumor) on 3DIRCADb, and 96.6% (liver) and 85.7% (tumor) on ATLAS. These results highlight the significance of T-MPEDNet in addressing the challenges of liver and tumor segmentation in CT-MRI, establishing T-MPEDNet as a robust solution for clinical applications. © 2025 Elsevier Ltd