Explainable Split-Learning-Based Framework for Accurate Pulmonary Nodule Classification

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• Lung cancer is the leading cause of cancer death globally.
• Computer-Aided Diagnosis Systems (CADx) improve pulmonary nodule classification but face data limitations and privacy concerns.
• Existing CADx systems struggle with false negatives and false positives.

Purpose of the Study:

• To develop a privacy-preserving, collaborative feature extraction approach for efficient and diverse lung nodule classification.
• To reduce false positives and false negatives in lung cancer diagnosis using Computed Tomography (CT) scans.
• To enhance the trustworthiness and generalizability of AI models in medical diagnostics.

Main Methods:

• A novel explainable feature-based split learning approach was proposed.
• Utilized a split ResNet-50 architecture for client-side feature extraction.
• Employed a hybrid 2D-CNN with an attention mechanism on the server-side for classification.
• Evaluated using ablation studies (ConvNeXt-Tiny, EfficientNetB0) and external datasets.
• Assessed trustworthiness with Local Interpretable Model-agnostic Explanations (LIME) and Grad-CAM.

Main Results:

• Achieved 99.38% accuracy and F1-score with 1.23% false negatives and 0% false positives on the primary dataset.
• Demonstrated robustness on unseen datasets, yielding 99.28% accuracy (1.24% FN, 0% FP) on one and 95.74% accuracy (7.07% FN, 1.41% FP) on another.
• Confirmed data diversity and privacy preservation through evaluations.

Conclusions:

• The proposed explainable split learning approach effectively enhances lung nodule classification accuracy and efficiency.
• The method ensures data privacy and diversity while significantly reducing diagnostic errors.
• The model exhibits strong generalizability and trustworthiness, making it a promising tool for real-world clinical application.