Transformer-Driven Explainable Deep Learning with Quantitative Attribution Validation for Liver Tumor Detection

The identification of liver tumors on computed tomography (CT) scans is hindered by myriad factors, including tumor heterogeneity, anatomical variability, and the limited interpretability of deep learning models in clinical settings. The present research introduces a deep learning-based framework, referred to as the ‘form of the Transformer’, in combination with Global Context (GC) fused with Transformer (Tf) and the Quantitative Attribution (QA) module, for a first reliable, explainable liver tumor detection framework. Moving away from traditional opaque classification systems, this framework uses gradient-based attribution with a localization module and evaluates its spatial alignment with tumor annotations without requiring segmentation supervision during model training. The framework accounts for long-range spacing and leverages Tf-Encoders, which substantially improve the system’s tumor-detection performance. Integrating the Attribution, this framework significantly enhances Qualitative Evidence (QE) in clinical settings. The experimental study has shown strong classification performance with the following metrics: accuracy 96.9%, precision 96.2%, recall 95.8%, F1-score 96.0%, area under the receiver operating characteristic curve 97.6%, and Matthews correlation coefficient 0.93. The classification-based localization of the system achieves an Intersection over Union (IoU) of 71.6% and a Dice coefficient of 83.5%, underscoring the alignment of tumor regions with their attributions. The results indicated significant improvements over existing CNN- and TF-based systems.