Non-small cell lung cancer detection through knowledge distillation approach with teaching assistant

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• Non-small cell lung cancer (NSCLC) accounts for 85% of lung cancer cases and has a slower metastasis rate than small cell lung cancer.
• Accurate and efficient classification of NSCLC is crucial for timely diagnosis and treatment planning.

Purpose of the Study:

• To develop and evaluate a knowledge distillation framework for improved NSCLC classification using CT scan images.
• To enhance the performance and transparency of deep learning models in lung cancer detection.

Main Methods:

• Employed a three-stage knowledge distillation technique with teacher, teaching assistant (TA), and student deep learning models (CNN, VGG19, ResNet152v2, Swin, CCT, ViT).
• Utilized cost-sensitive learning and hyperparameter tuning (alpha, temperature) for optimal model performance.
• Applied explainable AI (Shapley values, partition explainer) for model transparency and developed a user-friendly web application for classification.

Main Results:

• The TA (ResNet152) and student (CNN) models achieved test accuracies of 90.99% and 94.53%, respectively, with optimal hyperparameters (alpha=0.7, temperature=7).
• The TA framework significantly improved the student model's performance.
• The knowledge distillation technique reduced trainable parameters and training time, making it suitable for memory-constrained edge devices.

Conclusions:

• The proposed knowledge distillation framework effectively enhances NSCLC classification accuracy and efficiency.
• Explainable AI methods provide transparency into the deep learning model's decision-making process.
• The developed system offers a practical and efficient solution for lung cancer detection, applicable to edge computing environments.