The Effectiveness of Self-supervised Pre-training for Multi-modal Endometriosis Classification*†
Summary
Self-supervised pre-training significantly enhances endometriosis classification from medical images. This approach overcomes challenges like small datasets and subtle lesions, improving diagnostic accuracy by up to 31%.
Area of Science:
- Medical imaging analysis
- Artificial intelligence in healthcare
- Women's health diagnostics
Background:
- Endometriosis affects 5-10% of women globally, necessitating early detection and treatment.
- Current imaging-based diagnosis faces challenges due to limited expert clinicians and subtle, unlocalized lesions, leading to small, difficult training datasets and overfitting in classifiers.
- Self-supervised pre-training methods offer a promising solution for improving generalization in machine learning models.
Purpose of the Study:
- To evaluate the effectiveness of modern self-supervised pre-training techniques for endometriosis classification using multi-modal imaging data.
- To introduce a novel masking image modeling self-supervised pre-training method for 3D multi-modal medical imaging.
- To present the first endometriosis classifier fine-tuned from a self-supervised pre-trained model using multi-modal (T1 and T2) MRI data.
Main Methods:
- Utilized self-supervised pre-training techniques, including a novel 3D masking image modeling approach.
- Developed and fine-tuned an endometriosis classifier using multi-modal (T1 and T2) MRI data.
- Compared the performance of the fine-tuned classifier against models trained from scratch.
Main Results:
- Self-supervised pre-training demonstrated significant improvements in endometriosis classification accuracy.
- The proposed method achieved up to a 31% increase in classification performance compared to training from scratch.
- The developed classifier effectively handles challenges associated with small datasets and difficult-to-detect lesions.
Conclusions:
- Self-supervised pre-training is a highly effective strategy for improving endometriosis classification from multi-modal MRI.
- The novel 3D masking image modeling method shows promise for medical imaging applications.
- This work establishes a new benchmark for AI-driven endometriosis diagnosis using multi-modal imaging.
