Correlation Routing Network for Explainable Lesion Classification in Multi-Parametric Liver MRI.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• Liver tumor diagnosis is crucial in abdominal imaging.
• Magnetic Resonance Imaging (MRI) offers unique advantages but faces challenges in automated classification.
• Existing research often focuses on CT and ultrasound, with fewer studies on MRI for focal liver lesions (FLL).

Purpose of the Study:

• To propose an explainable AI model for classifying liver lesions using multi-sequence MRI.
• To improve the accuracy and clinical accountability of automated liver tumor diagnosis.
• To address the technical complexity and dataset curation issues in liver MRI analysis.

Main Methods:

• Developed a Correlation Routing Network (CRN) utilizing 10 MRI sequences.
• CRN incorporates encoding branches, correlation routing/relay modules, and self-attention mechanisms.
• The model predicts lesion types (HCC, Cholangioma, Metastasis, Hemangioma, FNH, Cyst) and detailed imaging features.

Main Results:

• Achieved 97.2% accuracy for malignant-benign classification.
• Reached 88% accuracy for six-class lesion classification.
• Obtained an average imaging feature accuracy of 84.9%, outperforming CNN and transformer models.
• Identified signal relations for quantitative explainability.

Conclusions:

• The CRN model demonstrates high accuracy and explainability in liver MRI analysis.
• The approach enhances clinical accountability through detailed feature prediction.
• This work offers insights into multimodal lesion classification and AI model explainability in medical imaging.