Automated Detection of Small Low-Contrast Hyper-Reflective Foci in Optical Coherence Tomography Using a Two-Stage Deep Learning and Gradient Boosting Pipeline

Hyperreflective foci (HRF) are small, dot-like structures observable in optical coherence tomography (OCT) scans and are increasingly recognized as important biomarkers of neuro-retinal inflammation in diabetic retinopathy (DR). Due to their minuscule size ( 3×3 pixels) and low contrast, accurate segmentation and quantification of HRFs remain highly challenging in clinical practice. In this study, we propose a novel deep learning pipeline designed specifically for the fine-grained segmentation and counting of HRFs in OCT images. Our approach begins with a frequencydomain and spatial-domain preprocessing pipeline that enhances HRF visibility through Fast Fourier Transform (FFT) filtering, Gaussian blur, non-local means denoising, and Sobel edge enhancement. This preprocessing emphasizes the frequency characteristics and boundary contrast of HRFs, improving their separability from background noise. A custom residual U-Net architecture is then trained to generate detailed, pixel-level segmentation maps capable of capturing the subtle and localized features of HRFs. Finally, a post-processing algorithm based on frequency-selective contrast filtering is applied to suppress false positives and refine HRF detection. Evaluated on a dataset of 98 annotated OCT scans using leave-one-out cross-validation, our best-performing model (ResUNet-18) achieved a recall of 0.795 and an F1 score of0.658. These results demonstrate the potential of our method for precise and fine-scale segmentation of HRFs, supporting reliable and scalable biomarker quantification for clinical decision-making in DR management.