Automatic selection of optical coherence tomography images for prognostic prediction models in age-related macular...

Age-related macular degeneration (AMD) is a leading cause of blindness. Current standard treatments require frequent intravitreal injections and entail high costs, placing a heavy burden on patients and healthcare providers. These challenges often lead to treatment discontinuation or overtreatment, highlighting the need for personalized AMD management. Accurate early prediction of long-term treatment outcomes is critical for optimizing these strategies. However, most existing prognostic models rely heavily on manual image selection by ophthalmologists. This labor-intensive process, which requires carefully selecting suitable images from large volumes of electronic medical record (EMR) data, significantly hinders real-world implementation. This study proposes an automated deep learning framework to select appropriate images from extensive EMR-stored optical coherence tomography (OCT) reports, thereby reducing the reliance on manual curation.

We developed a vision transformer (ViT)-based architecture to perform automated image selection. The model integrates features from the fundus infrared reflectance (IR) images and the corresponding OCT images presented in the clinical reports.

Compared to related works using a single OCT image input, the proposed method outperformed to the baseline. At training the pretrained ViT framework achieved an overall accuracy of 89% in identifying suitable images. Furthermore, using the images selected by our method improved the downstream prognostic prediction accuracy by an average of 3.1 percentage points. Confidence scores showed a statistically significant difference between the proposed method and the baseline (p = 0.026). The 95% confidence interval (CI) of the performance difference was [0.002, 0.066].

These results demonstrate the feasibility of an automated module capable of reliably identifying suitable images for AMD prognosis. By streamlining image selection workflows, this approach can enhance clinical efficiency, support the accurate estimation of long-term treatment effects, and facilitate treatment planning.