Multi-task OCT for Octascope

This dataset supports the development of Octascope, a high-performance foundation model for optical coherence tomography (OCT) image analysis. The dataset comprises 449,000+ OCT images spanning multiple tissue types, including retina, kidney, and abdominal tissues, curated from publicly available sources and prior research efforts. These images were used in a curriculum learning strategy to pre-train convolutional neural networks on both general (ImageNet) and domain-specific OCT features. The resulting foundation model demonstrated improved predictive accuracy and computational efficiency in two downstream classification tasks: spinal tissue identification for epidural needle guidance and multi-label retinal disease diagnosis. Compared to prior models like RETFound, Octascope achieved up to 9.13% higher accuracy while maintaining 4–5× faster inference speeds. This dataset enables reproducible research in OCT imaging and serves as a benchmark for evaluating transfer learning strategies and model generalization across diverse medical imaging tasks.