Perception of retinal images: Can artificial intelligence help us discover new diagnostic features?
Bibliographic record
Abstract
Medical images are a rich source of information regarding health. Diagnosticians are trained to sift through them to detect subtle signs of pathological processes, and to ignore vast variations unrelated to pathology. Retinal images are routinely used in the diagnosis and management of ocular diseases. Might there be signs of pathology in a retinal image, beyond eye diseases, that are hiding in plain sight, but currently overlooked? Convolutional neural networks (CNN) trained on retinal fundus images can classify patient sex, a trait that is invisible to the diagnostician (e.g. ophthalmologist) in this modality. Recent work in the interpretation of a CNN model trained for sex classification has elucidated features within fundus images that were relevant to this task (Delavari et al., 2022). Using patient sex as a case study, we investigated whether human observers can be trained to recognize “invisible” patient traits from fundoscopic images. We examined a group of diagnosticians (Expert, N=23) and a comparison group (Non-expert, N=31). In the pre-training phase, baseline sex recognition was assessed via a 2-alternative forced-choice (2-AFC) task without feedback. This was followed by a training phase and practice trials with feedback. Finally, a post-training 2-AFC sex recognition test and a novel object memory test (NOMT) (Richler et. al., 2017) to assess general object recognition ability were completed. Results for the pre-test are consistent with chance-level performance, M=52% for Experts, and M=52% for Non-experts, as expected. Post-test performance was significantly improved for Experts with M=66.1% (d= 2.38, p<<0.01) and for Non-experts M=66.2% (d=1.67, p<<0.01). Performance on the NOMT test was not related to improvement in fundus classification. Together, these results demonstrate that diagnosticians can be trained to recognize novel retinal features suggested by artificial intelligence. Future work with this approach can be extended to discover signs of systemic and neurodegenerative disease in retinal images.
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 machine prediction
Teacher imitationNot calibrated prevalence, not ground truth. Human validation pending. The Gemma side is a direct model label for every work in the frame, read from the title-only record. The Codex side is a classifier learned from the 10,348 direct Codex labels and calibrated to design-weighted sample rates; fields without enough sample support carry no Codex call. Candidate is the union of the two sides; consensus is their intersection. These outputs are machine_predicted_unvalidated and are not human labels.
Distilled classifier scores by category (both heads)
| Category | Codex | Gemma |
|---|---|---|
| Metaresearch | 0.002 | 0.015 |
| Meta-epidemiology (narrow) | 0.001 | 0.000 |
| Meta-epidemiology (broad) | 0.001 | 0.001 |
| Bibliometrics | 0.001 | 0.000 |
| Science and technology studies | 0.000 | 0.002 |
| Scholarly communication | 0.003 | 0.005 |
| Open science | 0.001 | 0.001 |
| Research integrity | 0.001 | 0.002 |
| Insufficient payload (model declined to judge) | 0.003 | 0.001 |
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 source (direct Gemma or distilled Codex), 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".