Shape and margin-aware lung nodule classification in low-dose CT images via soft activation mapping

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Radiology

Background:

• Convolutional neural networks (CNNs) are used for lung nodule classification, but often lack clinical interpretation.
• Existing interpretation methods like Class Activation Mapping (CAM) and Gradient-based CAM (Grad-CAM) overlook fine-grained pathological features crucial for accurate lung nodule categorization in low-dose CT images.

Purpose of the Study:

• To develop advanced interpretation methods for CNNs in lung nodule classification.
• To enable the analysis of fine-grained lung nodule shape and margin (LNSM) features for enhanced diagnostic sensitivity and specificity.

Main Methods:

• Developed Soft Activation Mapping (SAM) to capture fine-grained, discrete features of lung nodules.
• Proposed a High-level Feature Enhancement Scheme (HESAM) integrating SAM with high-level convolutional features to localize LNSM features.
• Conducted experiments on the LIDC-IDRI dataset and a visually matching experiment with radiologists.

Main Results:

• SAM identified more discrete attention regions compared to existing methods.
• HESAM demonstrated state-of-the-art predictive performance in localizing LNSM features, significantly reducing false positive rates.
• A radiologist study confirmed the clinical applicability and increased confidence in the proposed method.

Conclusions:

• SAM and HESAM offer superior interpretation for CNN-based lung nodule classification by focusing on fine-grained shape and margin features.
• The proposed methods improve diagnostic accuracy and reduce false positives in low-dose CT interpretation.
• The integration of radiologist feedback validates the clinical potential of these advanced interpretation techniques.