Dual-model weight selection and self-knowledge distillation for medical image classification.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Deploying large-scale AI models in healthcare is hindered by computational resource constraints.
• Developing efficient, lightweight models for medical image classification is essential for practical clinical application.
• Existing methods often struggle to retain critical information in compact models.

Purpose of the Study:

• To propose a novel medical image classification method that integrates dual-model weight selection with self-knowledge distillation (SKD).
• To develop lightweight models that achieve performance comparable to large-scale models while maintaining computational efficiency.
• To overcome the limitations of conventional approaches in retaining critical information within compact models.

Main Methods:

• Employed a dual-model weight selection strategy to initialize two lightweight models using weights from a large pretrained model for effective knowledge transfer.
• Applied self-knowledge distillation (SKD) to selected models, enabling diverse initial weight configurations without significant computational overhead.
• Fine-tuned the models for specific medical image classification tasks.

Main Results:

• Demonstrated superior performance and robustness of the proposed method compared to existing approaches.
• Validated the effectiveness across diverse medical imaging datasets, including chest X-rays, lung CT scans, and brain MRI scans.
• Achieved comparable performance to large-scale models with significantly reduced computational requirements.

Conclusions:

• The proposed method effectively addresses computational constraints in medical AI deployment.
• Combining dual-model weight selection with SKD offers a robust and efficient solution for medical image classification.
• The approach shows significant potential for practical implementation in real-world clinical settings.