gms | German Medical Science

Age-Related Macular Degeneration (AMD) is one of the most common eye disorder leading to vision loss [1]. Fluid and drusenoid structures in the retina are primary indicators for disease staging. Optical Coherence Tomography (SD-OCT) scans are widely employed to analyze retinal structures, allowing a precise differentiation between dry and wet cases. However, conventional diagnostic techniques are time-consuming and frequently rely on subjective interpretations, which may lead to inconsistencies in classification. We further enhanced our results from [2], where we successfully segmented biomarkers like drusen and neovascularizations (CNV), and added pseudodrusen with 0.78 F1. The segmentation process, based on the deep learning model U-net enabled the precise identification and delineation of these biomarkers in OCT scans, thereby facilitating a better analysis of retinal structures.

Building upon the segmentation capabilities developed in our earlier research, the current study improved the segmentation of fluid and drusenoid structures on OCT scans by employing a U-Net-based model with a mixed vision transformer backbone, which outperformed our previous backbone. The model was trained on 1400 labeled OCT-scans (600 wet, 700 dry) with 4881 instances of drusen, 188 pseudodrusen, 2810 CNVs, and 120 PEDs. We further implemented a linear interpolation between OCT slices, that allows the generation of high-resolution 3D tensors, facilitating a more accurate assessment of the size and distribution of the previously segmented drusen and fluid. Consequently, our approach serves as a step towards explainable AI in AMD classification, shifting the focus from automatic classification of AMD stages to providing detailed quantitative information on drusen size and fluid presence, which are essential to accurately distinguish between early, intermediate and late AMD [3]. The 3D visualization of pathological regions in the interpolated volume images enables a better view of retinal structures. This visualization is expected to aid clinicians in understanding the model's decision-making process, enhancing ttrust in the AI-based classification system.

In future work, we intend to integrate fundus images into the linear interpolation to capture smaller structures such as minor drusen, which may be missed due to the wide gap between OCT slices. The incorporation of additional imaging modalities is expected to improve the overall performance and accuracy of our interpolation. The combination of precise biomarker segmentation and advanced interpolation techniques is expected to offer a more detailed and reliable assessment of the size and distribution of drusen, CNV, and PED. This, in turn, will allow clinicians to make better-informed decisions regarding the clinical stages of AMD, and guide the selection of appropriate treatment options for patients with this progressive eye disorder.

Our results demonstrate the successful visualization of fluid and larger drusen in the 3D interpolated volume images. However, a comprehensive evaluation must be conducted to validate the model's segmentation and interpolation performance and to compare it with conventional diagnostic techniques. This study contributes to the development of explainable AI-based AMD assessment tools, which should provide clinicians with relevant quantitative information on drusenoid structures and neovascular fluid, and reduce their workload by facilitating a more efficient computer-aided classification of AMD.

The authors declare that they have no competing interests.

The authors declare that an ethics committee vote is not required.