Data and code reproducibility package for Few-shot retinal disease diagnosis from OCTA projections: foundation-model embeddings versus hand-crafted features

This record presents the companion data and code for the article "Few-shot retinal disease diagnosis from OCTA projections: foundation-model embeddings versus hand-crafted features". The study compares foundation-model embeddings against the hand-crafted descriptors of biomedical engineering for retinal disease diagnosis from optical coherence tomography angiography (OCTA) en-face projections. It uses both fields of view of the public OCTA-500 dataset. The 3-mm cohort holds 200 subjects with disease classes of 29, 6 and 5, and the 6-mm cohort holds 203 subjects with disease classes of 43, 35, 14, 11 and 9. Four modelling arms were evaluated on fifteen identical subject-level folds per cohort: hand-crafted descriptors with classical learners, supervised transfer of ImageNet-pretrained networks, frozen and low-rank-adapted foundation-model embeddings, and few-shot classification of the rarest classes. A three-rung comparison separates the effect of class size from that of imaging. The record aims at making every number reported in the article justifiable, traceable and reproducible. It holds the result tables and fold-level outputs of both cohorts, together with the subject-level manifests that rebuild the exact study set from a fresh OCTA-500 download. Every analysis step is provided as an executed Jupyter notebook and as a plain script, with one environment and run record per run. The pre-registered protocol and its decision log are included, as is every manuscript figure with the code that generated it. Every fold metric was re-derived from the stored out-of-fold predictions, for all 405 folds of the 3-mm cohort and all 2,685 folds of the 6-mm cohort, and the reported numbers are checked in a register of 367 claims. 01_results/ (1,259 files): every result table and fold-level output of both cohorts, including the out-of-fold predictions, the resolution ablation and the bit-depth ablation.02_derived_data/ (5 files): the subject-level manifests, the curation record, the task map and the identifier list of the class-size comparison.03_code_notebooks/ (41 files): every analysis step as an executed Jupyter notebook, with its source, explanation and outputs.04_code_scripts/ (49 files): the same code as plain Python scripts, with the scripts that build the figures and audits.05_environment_and_run_records/ (26 files): one record per run, giving the interpreter, library versions, GPU, seed, timings and configuration.Experiment_documentation_v1.0

Authors

Institutions

Publication Details

Journal
Zenodo (CERN European Organization for Nuclear Research)
Published
2026-11-14
DOI
https://doi.org/10.5281/zenodo.24223782
Primary Topic
Retinal Imaging and Analysis
Type
article
Field-Weighted Citation Impact
0.00
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
article

Data and code reproducibility package for Few-shot retinal disease diagnosis from OCTA projections: foundation-model embeddings versus hand-crafted features

Ahmed BaniMustafa
Zenodo (CERN European Organization for Nuclear Research)
Retinal Imaging and Analysis
article

Data and code reproducibility package for Few-shot retinal disease diagnosis from OCTA projections: foundation-model embeddings versus hand-crafted features

Ahmed BaniMustafa
article en

Abstract

This record presents the companion data and code for the article "Few-shot retinal disease diagnosis from OCTA projections: foundation-model embeddings versus hand-crafted features". The study compares foundation-model embeddings against the hand-crafted descriptors of biomedical engineering for retinal disease diagnosis from optical coherence tomography angiography (OCTA) en-face projections. It uses both fields of view of the public OCTA-500 dataset. The 3-mm cohort holds 200 subjects with disease classes of 29, 6 and 5, and the 6-mm cohort holds 203 subjects with disease classes of 43, 35, 14, 11 and 9. Four modelling arms were evaluated on fifteen identical subject-level folds per cohort: hand-crafted descriptors with classical learners, supervised transfer of ImageNet-pretrained networks, frozen and low-rank-adapted foundation-model embeddings, and few-shot classification of the rarest classes. A three-rung comparison separates the effect of class size from that of imaging. The record aims at making every number reported in the article justifiable, traceable and reproducible. It holds the result tables and fold-level outputs of both cohorts, together with the subject-level manifests that rebuild the exact study set from a fresh OCTA-500 download. Every analysis step is provided as an executed Jupyter notebook and as a plain script, with one environment and run record per run. The pre-registered protocol and its decision log are included, as is every manuscript figure with the code that generated it. Every fold metric was re-derived from the stored out-of-fold predictions, for all 405 folds of the 3-mm cohort and all 2,685 folds of the 6-mm cohort, and the reported numbers are checked in a register of 367 claims. 01_results/ (1,259 files): every result table and fold-level output of both cohorts, including the out-of-fold predictions, the resolution ablation and the bit-depth ablation.02_derived_data/ (5 files): the subject-level manifests, the curation record, the task map and the identifier list of the class-size comparison.03_code_notebooks/ (41 files): every analysis step as an executed Jupyter notebook, with its source, explanation and outputs.04_code_scripts/ (49 files): the same code as plain Python scripts, with the scripts that build the figures and audits.05_environment_and_run_records/ (26 files): one record per run, giving the interpreter, library versions, GPU, seed, timings and configuration.Experiment_documentation_v1.0

Zenodo (CERN European Organization for Nuclear Research)
Higher Colleges of Technology (AE)
Retinal Imaging and Analysis
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon