%Aigaion2 BibTeX export from Idiap Publications
%Tuesday 11 August 2026 05:35:05 PM

@INPROCEEDINGS{vanRijn_BEYONDTHELASTFRAME_2026,
                      author = {van Rijn, Tymo and Pulvirenti, Roberto and Shiri, Mahshid and Meloni, Ilenia and Jimenez-del-Toro, Oscar and Tomasoni, Mattia and Hoogewoud, Florence and Kaisari, Eirini and Anjos, Andr{\'{e}}},
                    projects = {Idiap, IMAGIN-AIR},
         mainresearchprogram = {AI for Everyone},
  additionalresearchprograms = {AI for Life},
                       month = sep,
                       title = {Beyond the Last Frame: Temporal Modelling of Fluorescein Angiography for Hyperfluorescence Classification},
                   booktitle = {Ophthalmic Medical Image Analysis},
                      series = {Lecture Notes in Computer Science},
                        year = {2026},
                    abstract = {Fluorescein angiography (FA) captures the passage of dye through the retinal vasculature over time, and the resulting hyperfluorescence patterns are central to assessing retinal vascular and inflammatory disease. Yet existing automated grading methods reduce each examination to a single frame, discarding the temporal evolution that clinicians rely on to distinguish these patterns. 
We investigate whether temporal modelling improves automated hyperfluorescence grading beyond single-frame approaches by representing each FA examination as a sequence of frames. We partition each sequence into Early, Mid, and Late phases using a data-driven procedure based on linear discriminant analysis of frame embeddings. Frames from each phase are encoded by the RETFound-Green retinal foundation model and aggregated by an attention-pooled gated recurrent unit (GRU) to classify five hyperfluorescence types. 
To enable temporal analysis, we reconstruct and release per-frame acquisition times in the public AngioReport FA dataset.
Evaluated on held-out patients over five random seeds, our pipeline consistently improves discrimination over the state-of-the-art single-frame approach across all five hyperfluorescence types, raising the overall AUROC from 0.82 to 0.87, with the largest gains for types defined by their temporal evolution. Modelling the temporal dimension of FA, rather than a single frame, thus consistently improves grading, and our released timestamped dataset opens temporal FA analysis to the community. Code and the reconstructed timestamps are available at https://gitlab.idiap.ch/medai/software/paper/miccai-omia-beyond-the-last-frame.},
                         pdf = {https://publications.idiap.ch/attachments/papers/2026/vanRijn_BEYONDTHELASTFRAME_2026.pdf}
}