SC-MSDNet: Spatial-consistent multi-view self-distillation for retinal OCT classification.

Area of Science:

• Medical Imaging Analysis
• Computer Vision
• Machine Learning

Background:

• Current optical coherence tomography (OCT) classification models show high accuracy on benchmark datasets.
• However, these models struggle with generalization when imaging conditions differ across datasets.

Purpose of the Study:

• To introduce a novel spatial-consistent learning framework for OCT classification.
• To improve the generalization capability of OCT models across diverse datasets and scanners.

Main Methods:

• Developed SC-MSDNet, which enforces depth-wise feature coherence as an explicit optimization constraint.
• Utilized L_SC loss for cross-view consistency within horizontal feature bands and L_MV loss for prediction agreement across augmented inputs.
• Trained exclusively on the OCT2017 dataset and tested on external datasets without fine-tuning.

Main Results:

• Achieved 99.5% accuracy on OCT2017, 98.71% on the Retinal OCT Kaggle C8 dataset, and 96.37% on OCTDL.
• Demonstrated strong cross-dataset generalization without dataset-specific fine-tuning.
• Produced spatially localized attribution maps using only image-level labels.

Conclusions:

• Spatial-consistent learning offers a robust solution to the generalization problem in OCT classification.
• Embedding spatial priors into the loss function addresses model brittleness caused by varying acquisition protocols.
• This approach advances OCT image analysis by enabling reliable deployment across different institutions and scanners.