Lung nodule malignancy classification with weakly supervised explanation generation.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer-Aided Diagnosis

Background:

• Explainable AI (XAI) aims to provide insights for decision-making, but often requires fully annotated data, which is scarce in medical imaging.
• Accurate classification and explanation of lung nodules are crucial for early cancer detection.

Purpose of the Study:

• To develop an innovative approach for classifying lung nodules (benign/malignant) in CT volumes.
• To generate morphologically meaningful explanations for the classification decisions.
• To address the challenge of limited fully annotated data in the medical domain.

Main Methods:

• A deep learning architecture trained with a multi-phase regime.
• Full supervision for nodule class label (benign/malignant) learning.
• Weakly supervised learning for semantic attributes (e.g., nodule margin, sphericity, spiculation).

Main Results:

• Achieved 89.1% accuracy and 0.91 AUC on the LIDC-IDRI dataset, comparable to state-of-the-art fully supervised methods.
• Generated eight attribute scores as explanations from a smaller training set.
• Demonstrated robustness by correctly labeling 95% of false positive or random regions as benign.

Conclusions:

• The proposed approach effectively handles computer-aided diagnosis with sparse, fully annotated image data.
• This method offers a practical solution for integrating explainable AI in medical image analysis.
• The system provides both accurate classification and interpretable insights for lung nodule diagnosis.