Intra- and Inter-Slice Contrastive Learning for Point Supervised OCT Fluid Segmentation.

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Optical Coherence Tomography (OCT) fluid segmentation is vital for diagnosing and treating eye conditions.
• Current methods using pixel-wise annotated masks with Convolutional Neural Networks (CNNs) are effective but require extensive manual annotation.
• The annotation process for OCT images is costly, time-consuming, and demands specialized expertise.

Purpose of the Study:

• To develop an efficient OCT fluid segmentation method that minimizes the need for detailed annotations.
• To introduce a novel network, the Intra- and inter-Slice Contrastive Learning Network (ISCLNet), for OCT fluid segmentation using only point supervision.
• To leverage unlabeled OCT data through contrastive learning strategies to improve segmentation accuracy.

Main Methods:

• Proposed ISCLNet utilizes contrastive learning tasks to learn visual representations from unlabeled OCT data.
• An intra-slice contrastive strategy exploits fluid-background similarity and retinal layer-background dissimilarity within a single OCT slice.
• An inter-slice contrastive architecture learns similarities between adjacent OCT slices in a volume, combined with point supervision for end-to-end training.

Main Results:

• ISCLNet successfully segments fluid in OCT images using only point supervision.
• The method demonstrates performance that bridges the gap between fully-supervised and weakly-supervised segmentation techniques.
• Experiments on public datasets (AI Challenger, RETOUCH) show ISCLNet outperforms existing point-supervised segmentation methods.

Conclusions:

• ISCLNet offers a more efficient and less labor-intensive approach to OCT fluid segmentation.
• The proposed contrastive learning strategies effectively utilize unlabeled OCT data for improved segmentation.
• This work advances automated analysis of OCT images, aiding in clinical diagnosis and treatment planning.