Explainable Knowledge Distillation for On-Device Chest X-Ray Classification.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Science

Background:

• Automated multi-label chest X-ray (CXR) classification uses deep learning but often requires high computational resources.
• This limits the deployment of advanced diagnostic tools on resource-constrained devices.

Purpose of the Study:

• To develop a compact deep learning model for real-time multi-label CXR classification using knowledge distillation (KD).
• To enhance model interpretability through explainable artificial intelligence (XAI).

Main Methods:

• A knowledge distillation strategy was employed, transferring knowledge from larger teacher models (CNNs, Transformers) to a smaller student model.
• Explainable AI (XAI) techniques were utilized to provide visual explanations for the distilled model's decisions.

Main Results:

• The KD strategy significantly improved the performance of the compact student model across three benchmark CXR datasets (ChestX-ray14, CheXpert, PadChest).
• For example, using DenseNet161 as a teacher, the EEEA-Net-C2 student model achieved AUCs of 83.7%, 87.1%, and 88.7% respectively.
• The compact model has only 4.7 million parameters and a computational cost of 0.3 billion FLOPS.

Conclusions:

• Knowledge distillation offers a feasible solution for deploying efficient deep learning models for CXR analysis on hardware with limited computational capacity.
• The proposed method balances model performance and computational efficiency, making AI-powered CXR diagnosis more accessible.