The paper evaluates three architectures for multi-label retinal disease classification on a unified 39‑target dataset. A staged gated cascade, a flat sigmoid model, and a label‑correlation transformer are compared, all sharing the same backbone and training protocol. The label‑correlation model yields the largest performance gain, and blending it with the flat model under test‑time augmentation achieves the best results across most metrics.
arXiv:2603. 18846v3 Announce Type: replace-cross Abstract: Foundation models are used to extract transferable representations from large amounts of unlabeled data, typically via self-supervised learning (SSL).
arXiv:2608. 12035v1 Announce Type: cross Abstract: Deploying unsupervised domain adaptation (UDA) in clinical practice requires choosing which algorithm to use and which of its trained models to ship.
arXiv:2610.10335v1 Announce Type: new Abstract: Visual in-context learning (ICL), well suited to label-scarce medical imaging, uses support image-label pairs to demonstrate input-output mappings, whi...
The paper presents a method for adapting a single vision‑language model, Qwen3.5‑9B, to medical image parsing tasks by using task‑specific low‑rank adapters (LoRA) selected from the input question. The system employs a frozen 4‑bit backbone and customizes adapter ranks, prompts, and class rebalancing across six task families and seven imaging modalities, achieving strong performance on the FLARE 2026 Task 3 dataset and a hidden‑validation set. The authors provide code and demonstrate that task‑specific adapter configurations and image budgets improve performance across diverse medical imaging tasks.
The study evaluates self‑supervised pretraining for Prognostics and Health Management (PHM) across 47 public datasets, using a compact Transformer encoder trained with masked patch prediction. Transfer performance is highly task‑dependent: it dramatically improves fault‑classification on CWRU and HSG18 but offers no reliable benefit for remaining‑useful‑life estimation on CMAPSS. The authors also compare centralized versus simulated federated pretraining, finding that centralized training generally outperforms federated training under matched budgets, and that the intrinsic SSL loss does not reliably predict downstream performance.
Shortcut learning occurs when classifiers rely on spurious correlations instead of true diagnostic features. The authors argue that this is essentially a calibration issue: unconstrained training aligns shortcut groups with their prevalence, causing over‑confidence in one group and under‑confidence in another. They introduce two encoder‑agnostic methods—a regularizer and a post‑hoc recalibration—to equalize prevalence across shortcut groups, achieving significant performance gains on chest‑drain‑pneumothorax benchmarks with both fine‑tuned CNNs and frozen foundation‑model backbones.
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.
arXiv:2609.25850v1 Announce Type: new Abstract: Deep learning performance generally improves with increasing training data, yet this scaling is fundamentally constrained by annotation cost in large-s...
arXiv:2610.17573v1 Announce Type: cross Abstract: Self-supervised pretraining has expanded the pool of 3D medical foundation models, but published evaluations use different datasets, endpoints and fi...
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.
The study investigates when adding zero‑shot pseudo‑labels from vision‑language models (VLMs) to small expert‑labelled medical imaging datasets ceases to improve, and may even degrade, classifier performance. Across three benchmarks—PatchCamelyon, ISIC 2019, and NIH ChestX‑ray14—the crossover point occurs after roughly 20–50, 100, and 250–500 expert labels, respectively, with additional pseudo‑labels reducing macro‑AUC by up to 0.14 points. The analysis shows that the VLM’s labeling quality, rather than the student model size, limits performance, and a simple bias‑variance rule can predict the benefit of pseudo‑labels in most settings.
The paper introduces RetinaGroundBench, a new benchmark for visual grounding of retinal OCT images, featuring 815 B-scans annotated with 4,537 bounding boxes for 13 biomarkers across 8 diseases. It evaluates 27 multimodal large language models (MLLMs) and finds that visual grounding is largely unsolved, with the best model achieving only a macro F1 of 23.7%. The authors also present RetGrounder, an 8‑B parameter MLLM fine‑tuned for OCT, which raises grounding F1 to 33.9%, a 43% relative improvement.