IDF-Net: Interpretable Dynamic Fusion Network for Colorectal Cancer Diagnosis Using Cross-Modal Imaging.

Area of Science:

• Artificial Intelligence in Medicine
• Oncology
• Medical Imaging Analysis

Background:

• Colorectal cancer (CRC) remains a major global health concern, necessitating advanced diagnostic tools.
• Current AI diagnostic models often lack interpretability, hindering clinical integration.
• There is a critical need for early, accurate, and clinically understandable CRC diagnostic solutions.

Purpose of the Study:

• To develop an interpretable dynamic fusion framework (IDF-Net) for multimodal CRC diagnosis.
• To integrate endoscopy, computed tomography (CT), and histopathology data for enhanced diagnostic accuracy.
• To improve the clinical interpretability of AI models in cancer diagnostics.

Main Methods:

• Proposed IDF-Net, a framework using modality-specific encoders, adaptive gating, and cross-modal attention.
• Conducted stratified 5-fold cross-validation for robust performance assessment.
• Employed spatial heatmaps and SHapley Additive exPlanations for interpretability analysis.

Main Results:

• IDF-Net achieved state-of-the-art accuracy (0.920) and AUC (0.991), outperforming baseline models.
• Interpretability analysis showed strong alignment between model predictions and expert annotations.
• Ablation studies confirmed the efficacy of dynamic routing and cross-attention fusion components.

Conclusions:

• IDF-Net offers a novel, interpretable, and dynamically fused multimodal approach for CRC diagnosis.
• The framework demonstrates superior accuracy and significant potential for clinical translation.
• IDF-Net's adaptive design ensures robustness, even with missing CT data, supporting flexible clinical workflows.