A semantic fidelity interpretable-assisted decision model for lung nodule classification.

Area of Science:

• Artificial Intelligence in Medical Imaging
• Computer-Aided Diagnosis
• Lung Cancer Detection

Background:

• Early lung nodule diagnosis is critical for lung cancer treatment.
• Existing capsule network models offer interpretability but struggle with robust feature extraction in shallow networks.
• This limitation hinders overall model performance.

Purpose of the Study:

• To propose a semantic fidelity capsule encoding and interpretable (SFCEI) model for lung nodule multi-class classification.
• To enhance the feature extraction capabilities of shallow capsule networks.
• To improve the accuracy and interpretability of lung nodule classification models.

Main Methods:

• Developed a multilevel receptive field feature encoding block to capture multi-scale lung nodule features.
• Integrated these blocks into a residual code-and-decode attention layer for fine-grained context extraction.
• Formulated semantic fidelity lung nodule attribute capsule representations by combining multi-scale and contextual features.

Main Results:

• Achieved a classification accuracy of 94.17% on the LIDC-IDRI dataset.
• Demonstrated superior performance compared to existing advanced approaches for lung nodule malignancy score classification.
• Validated through stratified fivefold cross-validation.

Conclusions:

• The proposed SFCEI methodology effectively captures multi-scale and contextual features of lung nodules.
• Enhanced feature-drawing capabilities in shallow capsule networks improve malignancy score classification.
• The interpretable nature of the model boosts physician confidence in clinical decision-making.