The paper presents an explainable diabetic retinopathy classification framework that leverages vision foundation models—DINOv2, CLIP, and Vision Transformer—combined with various transfer learning techniques such as full fine‑tuning, linear probing, and Low‑Rank Adaptation (LoRA). Using the ODIR dataset for internal validation and the APTOS dataset for external testing, DINOv2‑LoRA achieved the best internal AUROC (0.758) while DINOv2 and ViT full fine‑tuning reached the highest external AUROC (0.920). Explainability was assessed with Grad‑CAM and HiResCAM against expert‑annotated lesion masks from IDRiD, using Dice, IoU, and Pointing Game metrics, confirming that model attention aligns with clinically relevant retinal lesions.
By Abhishek Verma, Anila Krishna, Abhishek Gajanan Bankar, Juan Miguel Lopez Alcaraz
CF‑NODE is a counterfactual test‑time adaptation framework designed for diabetic retinopathy grading across different cameras, clinics, and patient populations. It ranks internal spatial nodes by attention, feature magnitude, and similarity, then progressively attenuates the highest‑ranked nodes and calibrates predictions based on the stability of the expected grade response. Evaluated on five external datasets, CF‑NODE outperforms four existing methods, achieving higher AUC and macro‑F1 scores and providing more balanced recognition across severity levels.
By Sabarish G, Deepak Parashar
The paper introduces FOCUS, a cross‑dataset benchmark for retinal fundus models that evaluates vision‑only encoders, vision‑language dual encoders, and multimodal large language models across ten public datasets. It measures ranking performance, calibration, subgroup disparities, and image‑quality robustness for tasks such as diabetic retinopathy and glaucomatous optic neuropathy. Results show no single model family dominates, with vision‑only encoders ranking highest on average, while fine‑tuning improves in‑domain performance but varies in external transfer.
By David Restrepo, Chenwei Wu, Luis Filipe Nakayama, Miguel L. Martins, Stergios Christodoulidis, Maria Vakalopoulou, Enzo Ferrante
The study evaluates fundus-specific foundation models (FM) for detecting diabetic macular edema (DME) in retinal images. It compares two popular FM—RETFound and FLAIR—against a lightweight EfficientNet-B0 backbone across multiple datasets (IDRiD, MESSIDOR-2, and OCT-and-Eye-FundusImages). Results indicate that FM do not consistently outperform fine‑tuned CNNs; EfficientNet-B0 often matches or exceeds FM performance, with FLAIR being the most competitive FM.
By Franco Javier Arellano, Jos\'e Ignacio Orlando
arXiv:2609.23561v1 Announce Type: new
Abstract: Deploying efficient neural networks is essential in resource-constrained environments, yet compact models often sacrifice interpretability - a critical...
By Aleks Czufarow, Ihor Babin
The study evaluates offline response-based knowledge distillation for five-class diabetic retinopathy grading on the APTOS 2019 dataset. Five teacher architectures—ResNet‑50, MobileNetV3‑Large, and Swin Transformer‑Tiny—were distilled into a MobileNetV3‑Small student, exploring a grid of teacher, temperature, and hard‑label weight settings. The best configuration achieved a quadratic weighted kappa of 0.9093 (macro‑F1 0.6851) on a single seed, with an average kappa of 0.9024 across three seeds, while reducing parameters by 18.07× compared to the Swin‑T teacher.
By Mossa Manisa, Habiba Jahan Priya, A. K. M. Fakhrul Hossain
The paper introduces Gated Knowledge‑Guided Prototype Transfer (GProtoT), a training‑free module that enhances zero‑support medical vision‑language models by transferring visual residuals from supported classes to missing ones using VLM similarity and clinical ontology distances. GProtoT employs an evidence gate to scale transfer based on semantic relevance, reducing negative transfer while modestly improving recall and accuracy in realistic one‑shot trials on MESSIDOR. The method requires no gradients, validation data, or extra inference, offering a low‑compute, accountable adaptation path for incomplete clinical supervision.
By Sumaiya Fazal, Rizwan Qureshi, Laleh Behjat, Samira Ebrahimi Kahou
MVC-Bench is a new benchmark designed to evaluate the calibration of vision‑language models (VLMs) and medical VLMs (Medical‑VLMs) for medical image classification. It tests calibration across robustness to modality, backbone, and domain shift; effectiveness of calibration strategies and prompt‑tuning methods; and stability under prompt‑template and random‑seed variations. The benchmark includes eight backbones, three medical modalities (fundus imaging, histopathology, chest X‑ray), and compares post‑hoc, train‑time, and zero‑shot calibration approaches, reporting accuracy, Expected Calibration Error (ECE), Maximum Calibration Error (MCE), and Adaptive Calibration Error (ACE) over 1,638 experiments, while also proposing a Multi‑Class Margin (MCM) regularization technique that improves ECE in most settings.
By Ashshak Sharifdeen, Shihab Aaqil Ahamed, Ufaq Khan, Muhammad Akhtar Munir Sujair Ibrahim, Mohamed Rafeek Mareer Ahamed, Yutong Xie, Imran Razzak, Muhammad Haris Khan
The study examines how the choice of camera—Canon versus Nikon—affects image quality and AI performance in retinal screening for diabetic retinopathy. In a Brazilian cohort of 8,524 patients, Nikon images were more than twice as likely to be of inadequate quality compared with Canon, despite being cheaper. Using a frozen pretrained retinal foundation model, the authors evaluated 12 disease classification tasks across different resolutions and classifiers, finding that camera-related performance differences vary widely and are not consistently mitigated by higher resolution or classifier choice.
By Lisa Y. W. Tang
arXiv:2606.15910v3 Announce Type: replace
Abstract: An answer from a vision-language model (VLM) to a medical image question can be accepted without review only under a threshold bounding the error o...
By Reza Khanmohammadi, Kundan Thind, Mohammad M. Ghassemi
MVC-Bench is a calibration-focused benchmark for medical vision‑language models, evaluating how well these models express confidence across different modalities, backbones, and domain shifts. It tests robustness to modality, backbone, and domain changes, the effectiveness of calibration and prompt‑tuning strategies, and stability under prompt‑template and random‑seed variations. The benchmark includes 1638 experiments, reporting accuracy and Expected Calibration Error (ECE) along with other calibration metrics, and introduces a simple train‑time calibration method, Multi‑Class Margin (MCM) regularization, that achieves the lowest ECE in most settings.
arXiv:2607. 02569v1 Announce Type: cross Abstract: This paper presents a safety-centered empirical evaluation of uncertainty-aware last-layer adaptation for referable diabetic retinopathy screening using RETFound, a self-supervised vision-transformer retinal foundation model used here as a frozen feature encoder, and the public APTOS 2019 and DDR diabetic retinopathy fundus image datasets.
By Karim Mardhani