A self-supervised pre-training method for lesion segmentation in oral potentially malignant disorders.

Area of Science:

• Medical Imaging
• Computer Vision
• Artificial Intelligence

Background:

• Supervised training for oral potentially malignant disorder (OPMD) image segmentation demands extensive annotated data, which is often scarce, especially in remote areas with limited medical specialists.
• This scarcity poses a significant challenge for developing accurate and reliable OPMD detection systems.

Purpose of the Study:

• To propose and evaluate a self-guided pre-training method for OPMD image segmentation.
• To enhance the performance of OPMD lesion segmentation by automatically extracting features from unlabeled images, thereby reducing the reliance on expensive pixel-level annotations.

Main Methods:

• Utilized a large internal dataset (3,417 OPMD photographs) and two external datasets for validation.
• Developed a multiscale separable attention masked image model (MS-SAMIM) combined with a self-guided mask generation method.
• Incorporated teacher-student consistency and contrastive learning constraints to improve model robustness.

Main Results:

• Achieved high performance in OPMD lesion segmentation: 84.35% Dice, 74.65% IoU, 81.84% sensitivity, and 88.02% precision on the internal dataset.
• Demonstrated strong performance on external validation datasets, with Dice scores above 81% and IoU scores above 69%.
• Outperformed various advanced fully supervised and other self-supervised segmentation approaches.

Conclusions:

• The proposed self-guided pre-training strategy effectively reduces the dependency on pixel-level annotations for OPMD lesion segmentation from clinical photographs.
• The method shows promise for improving OPMD detection in resource-limited settings.
• Further prospective validation is recommended prior to clinical implementation.