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

Current OCT classification models achieve high benchmark accuracy but exhibit poor generalization when imaging conditions vary between datasets. We introduce spatial-consistent learning, where depth-wise feature coherence becomes an explicit optimization constraint rather than a post-hoc interpretation tool. SC-MSDNet partitions convolutional features into horizontal bands, enforcing cross-view consistency within each region through LSC loss while LMV ensures prediction agreement across original, contrast-enhanced, and frequency-transformed inputs. Trained exclusively on OCT2017 (4 classes, 84,484 images) and tested on two external datasets without fine-tuning, the method achieves 99.5% accuracy on OCT2017, 98.71% accuracy on the 4-class subset of Retinal OCT Kaggle C8 dataset, and 96.37% accuracy on OCTDL, suggesting potential for cross-dataset generalization across different scanners and disease categories. The approach requires only image-level labels yet produces spatially localized attribution maps. By embedding spatial priors directly into the loss function, this framework addresses the fundamental brittleness of data-driven models when deployed across institutions with different acquisition protocols. This differs from prior self-distillation or multi-view OCT methods, where spatial consistency is either absent or applied post-hoc, rather than embedded as a trainable constraint.