Entropy and distance-guided super self-ensembling for optic disc and cup segmentation

Area of Science:

• Medical Imaging
• Computer Vision
• Ophthalmology

Background:

• Accurate segmentation of the optic disc (OD) and optic cup (OC) is vital for monitoring glaucoma progression in the elderly.
• Convolutional neural networks (CNNs) are widely used for OD and OC segmentation, but domain shift issues limit accuracy across different datasets.
• Unsupervised domain adaptation (UDA) is a key approach to address the domain shift problem in medical image segmentation.

Purpose of the Study:

• To propose a novel unsupervised domain adaptation method, entropy and distance-guided super self-ensembling (EDSS), to improve OD and OC segmentation accuracy.
• To enhance the robustness and discriminative power of segmentation models in the context of UDA.
• To effectively guide the domain adaptation process using multi-information fusion.

Main Methods:

• Developed a super self-ensembling (SSE) framework combining two self-ensembling models to learn more discriminative image features.
• Introduced Gaussian noise exponential moving average (G-EMA) to improve the robustness of the self-ensembling framework.
• Implemented a multi-information fusion strategy (MFS) to guide and enhance the UDA process.

Main Results:

• The proposed EDSS method achieved superior performance compared to state-of-the-art UDA segmentation techniques.
• Achieved Dice scores of 0.8442, 0.8772, and 0.9006 on RIGA+ dataset sub-datasets.
• Obtained a Dice score of 0.9154 on the REFUGE dataset.

Conclusions:

• EDSS effectively addresses the domain shift problem in OD and OC segmentation.
• The proposed method demonstrates significant improvements in segmentation accuracy for glaucoma detection.
• EDSS shows strong potential for clinical application in ophthalmology for early glaucoma diagnosis and monitoring.