Semi-supervised OCT lesion segmentation via transformation-consistent with uncertainty and self-deep supervision

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Optical coherence tomography (OCT) is crucial for diagnosing retinal diseases, but manual lesion segmentation is labor-intensive.
• Accurate automatic segmentation is vital for monitoring disease progression and treatment efficacy.
• Current methods demand extensive pixel-wise annotations, hindering scalability.

Purpose of the Study:

• To develop a novel semi-supervised framework for segmenting lesions in OCT images, named TCUS (transformation-consistent with uncertainty and self-deep supervision).
• To address challenges of blurred lesion boundaries and unreliable predictions in unlabeled data.
• To enhance segmentation accuracy for diverse lesion sizes and shapes.

Main Methods:

• Proposed TCUS framework utilizing an uncertainty-guided transformation-consistent strategy for robust regularization.
• Employed a student-teacher network approach where the student learns from reliable uncertainty-informed targets from the teacher.
• Integrated self-deep supervision to leverage multi-scale information from both labeled and unlabeled OCT data.

Main Results:

• The uncertainty-guided strategy effectively handles blurred lesion areas and improves prediction reliability.
• Self-deep supervision enhanced the segmentation of lesions across various scales, significantly boosting Dice coefficient scores.
• TCUS demonstrated superior performance compared to existing state-of-the-art semi-supervised segmentation methods on two OCT datasets.

Conclusions:

• TCUS offers an effective semi-supervised solution for OCT lesion segmentation, reducing annotation burden.
• The combination of uncertainty guidance and self-deep supervision leads to improved accuracy and robustness.
• This framework holds significant potential for advancing the clinical application of OCT imaging in retinal disease management.