Skip to content

Trend Analysis — Foundation Models in Medicine (2025)

Generated 2026-06-16 from 43 digested 2025 papers.

Overview

Foundation models in medicine matured in 2025 from a collection of single-modality encoders into a contested design space spanning vision-only self-supervised backbones, medical vision-language models (VLMs), and unified generalist systems with pixel-level grounding. The year's central tension is specialist vs. generalist: domain-specific medical pretraining (RETFound, MedSAM, CT-FM) still wins on data efficiency and fine-grained tasks, but rapidly scaling general-purpose encoders (DINOv2/v3, SigLIP2, video LVMs) increasingly close the gap, and several benchmarks now show no single model dominates. Three forces shape almost every paper: the impossibility of centralizing patient data (driving federated, model-merging, and continual-learning approaches), the scarcity of labeled volumetric/3D data (driving SSL paradigms like JEPA and SimCLR), and the demand for clinical trust (driving grounding, uncertainty, robustness, and prospective validation). By year's end the frontier had shifted toward 3D/volumetric models trained on uncurated health-system data, unified perception-plus-reasoning VLMs, and rigorous benchmarking that interrogates how FMs are adapted and evaluated rather than chasing single-task SOTA.

Timeline

  • 2025-01: Native-3D vision-only CT encoders arrive — CT-FM (148k scans, intra-sample contrastive) and MEDFORM (CT+clinical-tabular contrastive) establish vision-centric, non-text 3D pretraining.
  • 2025-02: Privacy-driven composition emerges — MedForge (open-source-style LoRA merging), Med-LEGO (training-free model fusion), MedTok (graph-aware EHR tokenizer), and RELICT (memorization/replica auditing).
  • 2025-03: Generalist-via-adaptation peaks — continual-learning segmentation (CL-Net, 235 anatomies), unified ReID (MaMI), few-shot domain adaptation (MFM-DA), and the CUPCase finding that GPT-4o beats clinical LLMs on rare cases.
  • 2025-04: Self-supervised paradigm innovation — CheXWorld brings JEPA world-modeling to radiographs; multi-teacher distillation agglomerates heavy FMs into lightweight students.
  • 2025-05: Scaling laws and generalist-VLM adaptation — BioVFM-21M characterizes SSL scaling; MedBridge repurposes frozen general VLMs via MoE; XMedGPT adds grounding/uncertainty/prognosis.
  • 2025-06–07: Multimodal multi-disease vision FMs (MerMED-FM) and a turn toward trustworthiness — de-identification (DCM-DeID), latent-fragility auditing (LAPD).
  • 2025-08–09: Benchmarking-as-contribution and continual learning consolidate — generalist-vs-specialist studies (ocular), data scaling laws for radiology, UNICON/EWC-diffusion-replay, and Citrus-V unified grounding.
  • 2025-10: External-validity reckoning — CardioBench and ProstNFound+ (first prospective FM validation) test whether FMs survive "beyond the lab"; video models probed as zero-shot medical learners.
  • 2025-11: Health-system-scale 3D FMs (NeuroVFM/Vol-JEPA, 5.24M volumes) and unified grounded VLMs (UMind-VL) define the frontier; comparative encoder benchmarks proliferate.
  • 2025-12: Generalist 3D-via-2D adaptation (AnyMC3D), retrieval-guided continual learning (PRIMED), surgical FMs (LapFM), and FMs used as semantic priors inside other pipelines (PGMP).

How the field developed

Early 2025 (Jan–Feb): building blocks and the privacy constraint. The year opened with a deliberate move toward native-3D, vision-only pretraining. CT-FM argued that text supervision entangles and degrades fine-grained visual features, pretraining a 3D encoder on 148k CT scans via intra-sample contrastive learning; MEDFORM paired CT with clinical tabular data. Simultaneously, the recognition that patient data cannot be centralized spawned a cluster of composition methods: MedForge framed FM-building as asynchronous open-source contribution of LoRA modules, Med-LEGO merged specialist adapters training-free via SVD-LoRA, and RELICT addressed the flip side—auditing generative FMs for memorized patient replicas. MedTok extended the FM idea to EHRs by tokenizing medical codes with their ontology graph context.

Spring 2025 (Mar–May): adaptation, continual learning, and SSL paradigms. With base FMs available, the dominant question became how to adapt them without retraining or violating privacy. CL-Net used continual learning + Lottery-Ticket-pruned decoders to segment 235 whole-body anatomies across 36 partially-labeled datasets, beating both nnUNet ensembles and SAM-style FMs. MaMI unified cross-modality ReID with runtime-generated LoRA weights; MFM-DA achieved few-shot (≤10 image) domain adaptation; FCA tackled federated client heterogeneity with decomposed adapters. A parallel thread pushed SSL paradigm innovation: CheXWorld imported JEPA world-modeling to chest X-rays at 20× less compute than Rad-DINO, while multi-teacher distillation (MedSAM2 + RAD-DINO + MedCLIP → one student) compressed heavy FMs. BioVFM-21M provided the first systematic scaling-law study, finding data diversity matters more than raw volume and that scaling benefits are highly task-dependent. CUPCase delivered a sobering counterpoint: general GPT-4o outperformed dedicated clinical LLMs on real, atypical case reports.

Mid 2025 (Jun–Aug): multimodality, trust, and survey consolidation. MerMED-FM demonstrated a single vision-only SSL model across 7 modalities and 10+ specialties (0.935 mean AUROC). The agenda visibly broadened from accuracy to trustworthiness: XMedGPT added visual grounding, sentence-level uncertainty, and prognosis; DCM-DeID decoupled identity from diagnostic semantics for adjustable de-identification; LAPD exposed that clinical LLMs with high BERTScore silently flip diagnoses under realistic perturbations; Dyslexify mechanistically defended medical CLIP against typographic attacks. A comprehensive GenAI survey positioned foundation models as the convergent architecture across the clinical workflow.

Late 2025 (Sep–Dec): benchmarking rigor, external validity, and the 3D/grounded frontier. The field grew self-critical. The ocular generalist-vs-specialist study, radiology data-scaling-laws, CardioBench, the eight-encoder feature-quality comparison, and the CXR-Foundation-vs-MedImageInsight benchmark all reached a consistent verdict: no single model dominates, adaptation strategy is often more decisive than the backbone, and general-purpose encoders (DINOv3, SigLIP2) are now competitive. AnyMC3D crystallized this by showing properly-adapted frozen 2D FMs beat 3D-native medical FMs, while diagnosing three pervasive benchmarking pitfalls. In parallel, the frontier scaled: NeuroVFM trained Vol-JEPA on 5.24M uncurated clinical volumes ("health system learning"), beating GPT-5 on report generation. Unified perception+reasoning VLMs (Citrus-V, UMind-VL) merged pixel-level grounding with diagnosis in one model. Continual learning was upgraded with retrieval (PRIMED's 18M-entry RAG memory) and exemplar-free diffusion replay (EWC-DDPM). FMs also became components—MedDINOv3 as a semantic prior inside an artifact-reduction pipeline (PGMP)—and expanded to new domains (LapFM for surgery, ProstNFound+ delivering the first prospective clinical validation with no degradation five years later).

Current state & major clusters

  • Native 3D / volumetric SSL. The clearest frontier. NeuroVFM (Vol-JEPA, 5.24M volumes, health-system learning), CT-FM (intra-sample contrastive), and the SimCLR 3D brain-MRI model train annotation-free volumetric encoders. AnyMC3D contests this by showing adapted 2D FMs + slice aggregation can beat 3D-native models, making "is 3D pretraining worth it?" an open empirical fight.

  • Unified generalist VLMs with pixel grounding. Citrus-V (detection + segmentation + multimodal CoT), UMind-VL (grounded ultrasound perception + clinical reasoning via an LLM-conditioned mask decoder), and XMedGPT (grounding + uncertainty + prognosis) integrate localization and reasoning that earlier MedSAM/LLaVA-Med-style models kept separate.

  • Generalist-vs-specialist benchmarking. A dominant contribution type in itself: CardioBench, the ocular study (RETFound-DINOv2 vs DINOv3), radiology data-scaling-laws (MI2/CLIP vs RAD-DINO/DINOv2), the eight-encoder feature-quality comparison, and CXR-Foundation-vs-MedImageInsight. Recurring findings: domain pretraining helps data efficiency, general scaling closes the gap, and retrieval/adaptation choices often matter more than the backbone.

  • Privacy-preserving composition & continual learning. MedForge and Med-LEGO (LoRA merging), FCA (federated adapters), and a strong CL line—UNICON (task/modality/anatomy expansion), CL-Net, EWC-diffusion-replay (exemplar-free), and PRIMED (retrieval-guided)—all assume data cannot be pooled and models must grow incrementally.

  • Parameter-efficient adaptation of frozen FMs. LoRA/adapters/MoE dominate downstream use: MedBridge (MoE over frozen VLMs), AnyMC3D (per-task LoRA), MedGemma-MSK and MedImageInsight-thoracic (frozen encoder + light head), ProstNFound+ (adapter-tuned MedSAM).

  • Trust, safety, and auditing. LAPD (latent diagnostic fragility), Dyslexify (typographic-attack defense), RELICT (memorization detection), DCM-DeID (de-identification), and PGMP's use of a frozen FM to suppress hallucination.

  • Domain expansion. Beyond CXR/CT/retina into surgery (LapFM), ultrasound (UMind-VL, ProstNFound+), echocardiography (CardioBench), dental CBCT (PGMP), and video/temporal data (zero-shot LVM study).

Open problems

  • Does native-3D pretraining beat adapted-2D? NeuroVFM/CT-FM vs AnyMC3D directly conflict; the cost-benefit of volumetric SSL versus 2D-FM + slice aggregation is unresolved.
  • Specialist vs. generalist convergence point. As DINOv3/SigLIP2 close gaps, it's unclear which modalities still justify domain-specific pretraining and at what data scale the crossover happens.
  • Benchmarking validity. AnyMC3D's three pitfalls (data-regime bias, suboptimal adaptation, narrow task coverage) imply much prior FM evaluation is unreliable; absolute clinical performance in realistic (not few-shot) regimes is often still inadequate.
  • External validity and prospective evidence. ProstNFound+ is a rare prospective study; almost all results remain retrospective, and cross-site/temporal generalization is largely unverified.
  • Latent robustness vs. surface metrics. LAPD shows benchmark scores hide diagnostic instability under clinically realistic perturbations; no standard robustness metric is adopted.
  • Catastrophic forgetting under privacy constraints. Continual learning still trades off intra- vs cross-domain memory; replay without exemplars (diffusion, retrieval) is promising but unproven at scale.
  • Hallucination and reproducibility in generative/diffusion FMs. Diagnostic reproducibility conflicts with stochastic sampling (PGMP), and memorization/privacy leakage (RELICT) lacks standardized QA.
  • Text supervision's role. Vision-only (CT-FM, MerMED-FM, NeuroVFM) vs vision-language pretraining show no consistent winner; whether reports help or entangle features is contested.
  • Fine-grained spatial tasks. Pretrained features remain insufficient for localizing subtle pathology (pneumothorax, lines & tubes) without heavy fine-tuning; supervised baselines (SegFormer, nnUNet) still compete.

Predicted next steps

  • Health-system-scale 3D FMs become the default for CT/MRI. Expect NeuroVFM-style Vol-JEPA pretraining on millions of uncurated, annotation-free clinical volumes to be replicated for chest/abdominal CT and whole-body MRI, with open releases challenging 2D-adaptation approaches head-to-head on standardized 3D benchmarks.
  • Benchmark standardization will be enforced. Following AnyMC3D/CardioBench, new FM papers will increasingly be required to report full-data (not just few-shot) regimes, multiple adaptation strategies, and multi-task/multi-modality coverage; "frozen + mean-pool" single-number evaluations will be treated as inadequate.
  • Convergence on adapted general-purpose encoders for many 2D tasks. As DINOv3/SigLIP2 keep closing gaps, more clinical 2D applications will ship a frozen general FM + lightweight LoRA/MoE rather than train a domain-specific backbone—reserving custom pretraining for 3D, ultrasound, and surgery where general models fail.
  • Grounded perception+reasoning VLMs spread to every modality. The Citrus-V/UMind-VL pattern (LLM-conditioned mask decoder + diagnosis in one model) will appear for CT, pathology, and dermatology, with grounding increasingly used as the mechanism for verifiable, hallucination-resistant outputs.
  • Retrieval-augmented and exemplar-free continual learning consolidate. PRIMED-style RAG memory and EWC-diffusion replay will merge into standard privacy-preserving update pipelines, displacing exemplar replay; expect multi-institution continual benchmarks (MGTIL-like) to become routine.
  • More prospective and external-validation studies. ProstNFound+ will be cited as the template; reviewers and journals will push for prospective or at least multi-site temporal validation before clinical claims, and "no degradation over time/site" becomes a headline metric.
  • FMs as priors/components, not just end models. Expect growth in using frozen medical FMs (MedDINOv3-style) as semantic-consistency or anti-hallucination regularizers inside reconstruction, artifact-reduction, de-identification, and generation pipelines.
  • Robustness metrics enter evaluation suites. LAPD-style latent-flip and adversarial (typographic) robustness checks will be increasingly bundled into FM benchmarks as trust becomes a gating requirement for deployment.

Key papers

  • Vision Foundation Models for Computed Tomography (CT-FM) (2025-01, Arxiv) — established native-3D, vision-only CT pretraining via intra-sample contrastive learning, arguing text supervision degrades fine-grained visual features.
  • MedForge (2025-02, Arxiv) — reframed FM-building as asynchronous open-source LoRA contribution, a template for privacy-preserving, decentralized model composition.
  • CL-Net (2025-03, Arxiv) — continual-learning segmentation of 235 whole-body anatomies across 36 partially-labeled datasets, beating nnUNet ensembles and SAM-style FMs at 55.5 MB.
  • CUPCase (2025-03, AAAI 2025) — real-world rare-case benchmark showing general GPT-4o beats dedicated clinical LLMs, puncturing exam-benchmark optimism.
  • CheXWorld (2025-04, CVPR 2025) — first JEPA world-model for radiographs, matching/beating large FMs with 20× less compute.
  • BioVFM-21M (2025-05, Arxiv) — first systematic SSL scaling-law study for medical vision; data diversity > raw volume, benefits are task-dependent.
  • MedBridge (2025-05, Arxiv) — repurposes frozen general-purpose VLMs for multi-label diagnosis via query injection + MoE, exemplifying adapt-don't-retrain.
  • MerMED-FM (2025-06, Arxiv) — vision-only multimodal/multi-disease FM (7 modalities, 0.935 mean AUROC) via memory-augmented teacher-student SSL.
  • Embeddings to Diagnosis (LAPD) (2025-07, Arxiv) — showed clinical LLMs flip diagnoses under realistic perturbations while surface metrics stay high, defining latent-robustness as a problem.
  • Citrus-V (2025-09, Arxiv) — unified medical FM merging detection, segmentation, and multimodal CoT reasoning with pixel-level grounding.
  • Data Scaling Laws for Radiology Foundation Models (2025-09, Arxiv) — controlled CLIP-vs-DINOv2 comparison; CLIP scales ~3× faster on findings, DINO dominates lines & tubes, motivating center-specific continual pretraining.
  • CardioBench (2025-10, Arxiv) — standardized echo benchmark revealing no model dominates and general encoders (SigLIP2, DINOv3) are competitive.
  • ProstNFound+ (2025-10, Arxiv) — first prospective clinical validation of an adapted medical FM, showing no degradation at a new site five years later.
  • Health system learning achieves generalist neuroimaging models (NeuroVFM) (2025-11, Arxiv) — Vol-JEPA on 5.24M uncurated clinical volumes; annotation-free 3D FM beating GPT-5 on report generation, defining "health system learning."
  • UMind-VL (2025-11, Arxiv) — first ultrasound FM unifying pixel-level grounded perception with clinical reasoning in one VLM.
  • Revisiting 2D Foundation Models for Scalable 3D Medical Image Classification (AnyMC3D) (2025-12, Arxiv) — showed adapted frozen 2D FMs beat 3D-native medical FMs and exposed three pervasive FM-benchmarking pitfalls.
  • PRIMED (2025-12, Arxiv) — retrieval-guided continual learning (18M-entry RAG memory) for generalist medical VLMs, advancing privacy-preserving, forgetting-resistant updates.