Curriculum learning for improved femur fracture classification: Scheduling data with prior knowledge and uncertainty

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Orthopedic Surgery

Background:

• Accurate classification of proximal femur fractures is vital for treatment and patient outcomes.
• The AO system provides a hierarchical classification based on fracture location and complexity.
• Current methods often struggle with limited, labeled datasets for training AI models.

Purpose of the Study:

• To develop an automated method for classifying proximal femur fractures using Convolutional Neural Networks (CNNs).
• To improve CNN performance in fracture classification by implementing a novel curriculum learning (CL) approach.
• To address challenges of limited data, class imbalance, and label noise in medical image datasets.

Main Methods:

• Proposed a curriculum learning (CL) framework integrating sample weighting, reordering, and subset sampling.
• Introduced two novel scoring functions: domain-specific prior knowledge and self-paced uncertainty.
• Experimented on a clinical dataset of proximal femur radiographs and the MNIST dataset.

Main Results:

• The CL approach significantly improved proximal femur fracture classification accuracy.
• The best performing CL method, reordering the training set by prior knowledge, achieved a 15% improvement.
• The unified CL formulation demonstrated benefits in limited data, class-imbalanced, and noisy label scenarios.

Conclusions:

• The proposed curriculum learning method enhances CNN performance for proximal femur fracture classification.
• This AI-driven approach can achieve classification accuracy comparable to experienced trauma surgeons.
• The CL framework offers a robust solution for improving AI model performance on challenging medical imaging datasets.