Using an adult retinal image analysis foundation model for retinopathy of prematurity staging: are there benefits?
Bibliographic record
Abstract
Retinopathy of prematurity (ROP) is a fibrovascular proliferative disease of the developing retina that can potentially lead to blindness in prematurely born infants. Detection of ROP in early stages is crucial for timely treatment. For ROP screening, specialized experts such as ophthalmologists are required, which may not be readily available. This has motivated the development of deep learning (DL) models to detect ROP and/or to stage its severity from color fundus photographs (CFPs) of the infant. The key challenge in the development of DL models for this task is the very limited availability of labeled training data. Recently, general data scarcity problems have been tackled in adult retinal image analysis with foundation models pre-trained on a huge number of (often) unlabeled images for auxiliary tasks, which are then fine-tuned for a specific downstream application. A representative example of a foundation model for adult retinal images is RETFound, which was trained on nearly one million unlabeled adult CFPs. Motivated by its public availability, this paper aims to investigate whether fine-tuning the adult RETFound model on a limited number of infant ROP images for disease severity staging leads to tangible performance benefits over fine-tuning generic deep learning-based image classifiers pre-trained on natural images from the ImageNet dataset. We perform extensive experiments on a large publicly available ROP dataset and surprisingly find that RETFound, despite having seen adult CFPs during pre-training, does not outperform the generic models. We believe that this finding highlights the unique challenges that pediatric medical image analysis presents and that the growing amount of adult foundation models may not help in alleviating data scarcity-related issues in pediatric applications.
Fetched live from OpenAlex and de-inverted. Abstracts are not stored in this database: the inverted indexes are 8.6 GB of the frame’s 9.3 GB of text, and the host has 13 GB free.
How this classification was reachedexpand
Full frame distilled prediction
Teacher imitationNot calibrated prevalence, not ground truth. Human validation pending. Learned from the 10,348 direct Codex labels and 10,348 direct Gemma labels. Candidate is the union of thresholded teacher heads; consensus is their intersection. These outputs are machine_predicted_unvalidated and are not human labels or direct frontier model labels.
Codex and Gemma teacher scores by category
| Category | Codex | Gemma |
|---|---|---|
| Metaresearch | 0.001 | 0.001 |
| Meta-epidemiology (narrow) | 0.000 | 0.000 |
| Meta-epidemiology (broad) | 0.001 | 0.000 |
| Bibliometrics | 0.000 | 0.001 |
| Science and technology studies | 0.000 | 0.000 |
| Scholarly communication | 0.000 | 0.000 |
| Open science | 0.000 | 0.000 |
| Research integrity | 0.000 | 0.000 |
| Insufficient payload (model declined to judge) | 0.000 | 0.000 |
Machine scores (provisional)
The two teacher heads of the student model, read on this work. A score orders the frame for review; it never asserts a category, and the validation status ships verbatim with every row.
Baseline scores from an immature model (maturity gate not passed, 7 training rounds). Scores rank; they never assert a category.
score_only:v0-immature-baseline · verbatim from the scoring run: score_only means the number may rank works, and no category label ships from itClassification
machine, unvalidatedMachine predicted; a candidate call from one teacher head, not a consensus.
How this classification was reached, model by model and score by score, is at the end of the page under "How this classification was reached".