Wednesday 05 March 2025
In recent years, scientists have made significant progress in developing techniques for weakly supervised segmentation of hyper-reflective foci in optical coherence tomography (OCT) images. These tiny specks, often appearing as small bright spots, are biomarkers for age-related macular degeneration (AMD), a leading cause of vision loss among older adults.
Researchers have traditionally relied on manual segmentation of these foci, which is both time-consuming and prone to errors. However, this approach becomes increasingly challenging with the increasing resolution and complexity of OCT images. To address this issue, scientists have turned to weakly supervised learning methods, where image-level annotations are used to train models for pixel-level segmentation.
One such approach involves using a compact convolutional transformer (CCT) model in conjunction with layer-wise relevance propagation (LRP). This combination enables the model to capture local spatial relationships and preserve high-resolution details, crucial for precise localization of these small structures. The CCT model is designed specifically for small datasets and can process images at full resolution.
In a recent study, researchers trained a CCT model on OCT images with varying levels of annotation intensity. They found that even a simple binary annotation scheme, where B-scans are labeled as either HRF-positive or -negative, was sufficient to achieve accurate segmentation results. This is in contrast to more time-consuming weak annotation styles, which provide additional information about the number and distribution of HRFs.
The researchers also developed an iterative inference strategy, where the model segments additional HRFs by masking out previously detected ones. This approach enables the model to refine its predictions over multiple iterations, leading to improved segmentation accuracy.
Another key innovation was the use of SAM 2, a foundation model capable of producing accurate segmentation masks based on minimal input prompts. The researchers developed a prompting strategy that leverages the relevance maps generated by LRP, allowing them to extract meaningful information from the complex OCT images.
The study’s results demonstrate the effectiveness of this approach in segmenting HRFs with high accuracy. By leveraging weak supervision and iterative inference, the model is able to detect even small HRFs amidst complex OCT images. This technology has significant potential for improving diagnostic accuracy and monitoring treatment efficacy in AMD patients.
Furthermore, this research highlights the importance of developing robust and efficient segmentation methods that can process high-resolution images at full resolution. As OCT imaging becomes increasingly widespread, such techniques will be essential for unlocking new insights into disease progression and response to therapy.
Cite this article: “Automated Segmentation of Hyper-Reflective Foci in Optical Coherence Tomography Images”, The Science Archive, 2025.
Optical Coherence Tomography, Age-Related Macular Degeneration, Weak Supervision, Segmentation, Convolutional Transformer, Layer-Wise Relevance Propagation, Iterative Inference, Sam 2, Foundation Model, Diagnostic Accuracy.
