[BibTeX] [RIS]
Green OCT AI: Evaluation of Diagnostic Performance and Energy Consumption of Ophthalmic Foundation Models, ARVO 2026, Denver, CO, USA 2026.
Tipo de publicação: Artigo
Citação:
Journal: IOVS, Investigative Ophthalmology & Visual Science
Volume: Vol. 67, n° 7
Ano: 2026
Mês: July
Resumo: Abstract Purpose : Foundation models (FMs) have gained notoriety in the field of ophthalmology; however, their energy efficiency and carbon footprint have rarely been reported. This research aims to assess the performance, energy consumption, and estimated CO2 emissions of ophthalmic FMs in the classification of ocular pathologies from OCT B-scans. Understanding the trade-off between performance and energy consumption is important for sustainable deployment in clinical pipelines. Methods : Three FMs, RETFound (ViT-L), VisionFM (ViT-B), and UrFound (ViT-B), were evaluated on the multi-class (n=7) OCTDL dataset (2063 images) for OCT B-scan ocular disease classification. Models were assessed across three downstream adaptation strategies: linear probing, k-nearest neighbours (k-NN), and full fine-tuning. Performance was analyzed using balanced accuracy (BAcc), AUROC, F1-score, recall, average precision (AP), and Cohen's kappa. Experiments were repeated across five random seeds. Computational efficiency and environmental impact were measured using CodeCarbon to track energy consumption (kWh) and CO2 emissions (kg) for each method, with all experiments conducted under identical settings. Results : Full fine-tuning achieved the highest BAcc, with UrFound reaching 0.91±0.022, followed by RETFound 0.89±0.015 and VisionFM 0.88±0.011(Table 1). Linear probing demonstrates competitive performance with VisionFM at 0.86±0.012, UrFound at 0.85±0.006, and RETFound at 0.80±0.006 while consuming approximately 100–500× less energy than fine-tuning (0.0002–0.0003 kWh vs. 0.04–0.14 kWh) (Figure 1). k-NN showed intermediate performance with minimal energy cost. While absolute energy values depend on hardware specifications, comparisons across methods remain valid and informative. Conclusions : FMs achieve strong classification performance but differ in energy demand. Our results demonstrate that linear probing achieves very good classification accuracy with a smaller carbon footprint. Fine-tuning delivers only 2-4% accuracy improvement with a much higher energy cost and higher CO2 emissions. These findings support parameter-efficient adaptation strategies for deploying FMs in ophthalmological workflows without compromising diagnostic utility or sustainability goals. This abstract was presented at the 2026 ARVO Annual Meeting, held in Denver, CO, May 3-7, 2026.
Palavras-chave:
Autores Rozhyna, Anastasiia
Müller, Henning
Santos, Rui
Adicionado por: []
Total mark: 0
Anexos
    Notas
      Tópicos