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

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• Liver tumor identification on CT scans faces challenges due to tumor heterogeneity, anatomical variations, and the opacity of deep learning models.
• Existing deep learning models often lack interpretability, hindering clinical adoption for liver tumor detection.

Purpose of the Study:

• To develop a reliable and explainable deep learning framework for liver tumor detection on computed tomography (CT) scans.
• To enhance the interpretability of deep learning models in clinical settings for liver tumor identification.

Main Methods:

• A novel deep learning framework combining Global Context (GC) fused with Transformer (Tf) and a Quantitative Attribution (QA) module was developed.
• The framework utilizes gradient-based attribution with a localization module, evaluating spatial alignment without requiring segmentation supervision during training.
• Transformer-Encoders were employed to capture long-range dependencies, improving tumor detection performance.

Main Results:

• The framework achieved high classification performance: 96.9% accuracy, 96.2% precision, 95.8% recall, 96.0% F1-score, 97.6% AUC, and 0.93 MCC.
• Classification-based localization yielded an Intersection over Union (IoU) of 71.6% and a Dice coefficient of 83.5%.
• Demonstrated significant performance improvements compared to existing Convolutional Neural Network (CNN) and Transformer-based systems.

Conclusions:

• The proposed deep learning framework offers a reliable and explainable solution for liver tumor detection on CT scans.
• The integration of attribution mechanisms enhances qualitative evidence, facilitating clinical decision-making.
• The framework shows superior performance and interpretability, paving the way for advanced AI applications in medical diagnostics.