Journal: Life

Article Title: HIRD-Net: An Explainable CNN-Based Framework with Attention Mechanism for Diabetic Retinopathy Diagnosis Using CLAHE-D-DoG Enhanced Fundus Images

doi: 10.3390/life15091411

Figure Lengend Snippet: Average performance comparison on the IDRiD-APTOS2019 dataset.

Article Snippet: To evaluate their framework, they used the Kaggle DR 2015 and APTOS2019 datasets.

Techniques: Comparison

Journal: medRxiv

Article Title: Act or Defer: Error-Controlled Decision Policies for Medical Foundation Models

doi: 10.64898/2026.02.23.26346927

Figure Lengend Snippet: a , Raw FM scores do not by themselves support safe clinical decisions: a high predicted risk may prompt treatment escalation, whereas a lower score may prompt monitoring, and both choices can be harmful when miscalibrated. b , Standard conformal prediction provides average (marginal) coverage, but the error rate can vary across predicted disease statuses. Multi-label prediction sets are often not actionable because they combine diagnoses with different management pathways. c , S trat CP performs error-controlled stratification: in the action arm , it selects predictions eligible for downstream action while controlling error among selected predictions within a user-specified budget (e.g., ≤5%) for each disease status; in the deferral arm , it returns prediction sets for deferred patients that achieve the target coverage (e.g., contain the true disease status for 95% of deferred patients), supporting calibrated follow-up. d , Ophthalmology: paired with RETFound FM on retinal images, S trat CP supports diabetic retinopathy staging, glaucoma staging, and multi-class eye-condition diagnosis by selecting patients in the action arm under the error budget and returning calibrated prediction sets for deferred patients; an optional utility module can reorder candidates within each set to provide clinically related differentials (e.g., shared follow-up actions). e , Neuro-oncology: paired with UNI FM on H&E whole-slide images, S trat CP supports biomarker prediction (IDH mutation status), CNS tumor subtyping, and time-to-mortality prognosis by selecting predictions in the action arm under the error budget and deferring the remainder to calibrated prediction sets (or one-sided lower prediction bounds for prognosis); for subtyping, guideline structure (e.g., WHO grade adjacency) can provide grade-coherent prediction sets.

Article Snippet: The raw datasets for the retinal diagnosis tasks are publicly available at the following sources: diabetic retinopathy diagnosis ( https://www.kaggle.com/competitions/aptos2019-blindness-detection/data ), glaucoma diagnosis ( https://dataverse.harvard.edu/dataset.xhtml?persistentId=doi:10.7910/DVN/1YRRAC ), and eye condition diagnosis ( https://zenodo.org/records/3477553 ).

Techniques: Biomarker Discovery, Mutagenesis

Journal: medRxiv

Article Title: Act or Defer: Error-Controlled Decision Policies for Medical Foundation Models

doi: 10.64898/2026.02.23.26346927

Figure Lengend Snippet: a , In the action arm, S trat CP maps FM predictions of patient profiles to candidate disease statuses and select predictions eligible for downstream action under a pre-specified error budget. Not-selected predictions are routed to the deferral arm, where S trat CP returns prediction sets calibrated to achieve the target coverage (e.g., the set contains the true disease status for 95% of deferred patients). b , S trat CP constructs a one-dimensional confidence score and calibrates a decision threshold using expert-labeled reference data, following a post-selection calibration procedure . Predictions with scores above the threshold are selected in the action arm (Methods). c , Guide to interpreting performance in terms of coverage and efficiency. A method is valid if it achieves coverage at least the target level and efficient if it selects more patients or slides while producing smaller prediction sets. A method is conservative if its coverage exceeds the target level but efficiency is low. A method is invalid and inefficient if its coverage is below the target level and efficiency is low. A method is aggressive if its coverage is below the target level even though efficiency is high. d , Diabetic retinopathy severity grading. From top to bottom: (i) coverage (1 – false discovery rate) within each predicted stage among selected patients; (ii) the number of selected predictions per predicted stage; (iii) marginal coverage over all held-out patients across target levels; and (iv) marginal prediction-set size over all patients across target levels. See Methods for metric definitions. e , Results for glaucoma diagnosis task. f , Results for eye condition diagnosis task. g , Results for IDH mutation status prediction task. h , Results for CNS tumor subtyping task.

Article Snippet: The raw datasets for the retinal diagnosis tasks are publicly available at the following sources: diabetic retinopathy diagnosis ( https://www.kaggle.com/competitions/aptos2019-blindness-detection/data ), glaucoma diagnosis ( https://dataverse.harvard.edu/dataset.xhtml?persistentId=doi:10.7910/DVN/1YRRAC ), and eye condition diagnosis ( https://zenodo.org/records/3477553 ).

Techniques: Construct, Labeling, Selection, Biomarker Discovery, Mutagenesis

Journal: medRxiv

Article Title: Act or Defer: Error-Controlled Decision Policies for Medical Foundation Models

doi: 10.64898/2026.02.23.26346927

Figure Lengend Snippet: a , Deferral arm: for a new deferred patient, S trat CP calibrates a prediction set using an expert-labeled reference group that would also be deferred under the same action-arm selection rule; the resulting sets achieve the target coverage level in the deferred group (e.g., contain the true disease status for 95% of deferred patients; Methods). b , Diabetic retinopathy: from top to bottom, deferred-set coverage (fraction of prediction sets containing the true disease status; target 95%), number of deferred patients, and mean set size at 95% coverage. S trat CP meets the 95% coverage target on deferred patients. c , Results for glaucoma diagnosis task. d , Results for eye condition diagnosis task. e , Results for IDH mutation status prediction task. f , Results for CNS tumor subtyping task.

Article Snippet: The raw datasets for the retinal diagnosis tasks are publicly available at the following sources: diabetic retinopathy diagnosis ( https://www.kaggle.com/competitions/aptos2019-blindness-detection/data ), glaucoma diagnosis ( https://dataverse.harvard.edu/dataset.xhtml?persistentId=doi:10.7910/DVN/1YRRAC ), and eye condition diagnosis ( https://zenodo.org/records/3477553 ).

Techniques: Labeling, Selection, Biomarker Discovery, Mutagenesis

Journal: medRxiv

Article Title: Act or Defer: Error-Controlled Decision Policies for Medical Foundation Models

doi: 10.64898/2026.02.23.26346927

Figure Lengend Snippet: a , Utility enhancement module. Given a CP utility graph encoding guideline relationships among disease statuses, S trat CP constructs a utility-informed ordering of candidate disease statuses. It starts from the status with the highest FM-predicted probability (e.g., Normal ), then adds high-probability alternatives that maximize utility with the statuses already selected (e.g., an adjacent stage such as Mild ); the final prediction set is the calibrated prefix of this ordering needed to achieve the target coverage in the deferred group (Methods). b , Diabetic retinopathy grading. We use an ordinal (chain) utility graph so that sets preferentially include adjacent stages; S trat CP orders labels by expanding along the severity scale. c , Fraction of prediction sets containing consecutive stages across target levels. Standard STRATCP is shown in blue, and utility-enhanced STRATCP is shown in orange. d , Coverage is maintained after utility enhancement (fraction of sets containing the true disease status) across target levels. e , Eye-condition classification. We construct a management-action utility graph whose edge weights quantify overlap in downstream management actions between disease-status pairs; actions are obtained by prompting an LLM (Methods). f , Mean overlap in management actions for prediction sets at matched set size; utility enhancement increases management coherence within sets. g , S trat CP maintains coverage across target levels. h , CNS tumor subtyping. We construct a WHO grade-based utility graph so that sets preferentially group subtypes within the same grade or adjacent grades (Methods). i , Fraction of prediction sets whose maximum within-set WHO grade difference is < 2 across set-size bins; utility enhancement increases grade coherence. j , S trat CP maintains coverage with the utility enhancement across target levels.

Article Snippet: The raw datasets for the retinal diagnosis tasks are publicly available at the following sources: diabetic retinopathy diagnosis ( https://www.kaggle.com/competitions/aptos2019-blindness-detection/data ), glaucoma diagnosis ( https://dataverse.harvard.edu/dataset.xhtml?persistentId=doi:10.7910/DVN/1YRRAC ), and eye condition diagnosis ( https://zenodo.org/records/3477553 ).

Techniques: Construct

Journal: International Journal of Ophthalmology

Article Title: Diabetic retinopathy identification based on multi-source-free domain adaptation

doi: 10.18240/ijo.2024.07.03

Figure Lengend Snippet: A–C: The loss of classification, PLG, and SCM for the RDR task; D–F: The loss of classification, PLG, and SCM for the NAD task. DR: Diabetic retinopathy; RDR: Referable diabetic retinopathy; SCM: Softmax-consistence minimization; PLG: Pseudo-label generator; NAD: Normal/abnormal detection.

Article Snippet: In , we observed that for the RDR task (DDR/EyePACS to APTOS2019), using only the source model in our SMPL resulted in an accuracy of 88.2% (89.67%).

Techniques: