Efficient urinary stone type prediction: a novel approach based on self-distillation.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Urology

Background:

• Urolithiasis necessitates accurate preoperative stone type identification for effective treatment.
• Current deep learning models for CT-based stone classification face challenges in size and computational efficiency for clinical use.

Purpose of the Study:

• To develop a non-invasive, computationally efficient deep learning approach for urinary stone type classification from CT images.
• To enhance the performance of lightweight models without relying on external teacher models or complex compression techniques.

Main Methods:

• Refinement of a self-distillation architecture incorporating feature fusion and the Coordinate Attention Module (CAM).
• Development of a method for effective knowledge transfer within lightweight models.
• Validation on proprietary and public datasets.

Main Results:

• Achieved 74.96% classification accuracy on a proprietary dataset, surpassing existing methods.
• Demonstrated superior performance and generalizability on two public datasets.
• The approach circumvents extra computational costs and performance reduction associated with model compression.

Conclusions:

• The proposed method offers an effective and efficient solution for non-invasive urinary stone classification using CT images.
• The model shows potential for clinical deployment, assisting in precise treatment planning and reducing patient discomfort.
• The approach's generalizability suggests applicability to other medical image processing tasks.