EVLF-FM: Explainable Med Vision-Language Model

• EVLF-FM is an explainable vision-language model for medicine that integrates multi-disease diagnosis, visual question answering, and grounding with pixel-level evidence.
• The model employs a dual encoder architecture—combining global disease-level and detailed pixel-level features—coupled with LoRA adapters and reinforcement learning for efficient fine-tuning.
• Empirical results show high diagnostic accuracy and visual grounding precision across various imaging modalities and clinical specialties, highlighting its practical clinical relevance.

EVLF-FM most specifically denotes the Explainable Vision-Language Foundation Model for Medicine, a multimodal medical vision-language foundation model optimized for multi-disease image diagnosis/classification, medical visual question answering, and medical visual grounding, with explicit emphasis on step-by-step reasoning and pixel-level visual evidence alignment (Bai et al., 29 Sep 2025). The term is nevertheless polysemous. In the dataset-distillation literature, the paper “EVLF: Early Vision-Language Fusion for Generative Dataset Distillation” does not define a component explicitly named “EVLF-FM”; in that context, “FM” is interpreted as the EVLF Fusion Module (Cai et al., 8 Mar 2026). In cardiology-oriented interpretations, the same string has also been mapped to distinct echocardiographic constructs, including full-volume intraventricular flow reconstruction and a generalized ejection-fraction quantity (Vixège et al., 2021, 1803.02825).

1. Nomenclature and scope

The acronym has been used in multiple unrelated technical settings. The dominant explicit usage is the medical foundation model introduced in 2025, but adjacent literature attaches the same string to different mechanisms or physiological quantities. Context is therefore essential.

Usage	Domain	Meaning
EVLF-FM	Medical AI	Explainable Vision-Language Foundation Model for Medicine (Bai et al., 29 Sep 2025)
EVLF Fusion Module	Dataset distillation	Lightweight cross-attention block inserted once between the encoder and the denoiser (Cai et al., 8 Mar 2026)
EVLF-FM as 3D-iVFM interpretation	Echocardiography	Full-volume, three-component intraventricular flow reconstruction from triplane color Doppler (Vixège et al., 2021)
EVLF-FM as TEBVF(t) interpretation	Echocardiography	Time-resolved generalization of ejection fraction based on LV myocardial segment volume changes (1803.02825)

Within medical AI, EVLF-FM is described as an early multi-disease VLM model with explainability and reasoning capabilities. Its stated objective is to unify broad diagnostic capability with fine-grain explainability, rather than remaining modality-specific or limited to black-box prediction (Bai et al., 29 Sep 2025).

2. Model architecture and representational design

EVLF-FM is organized around three main components: a disease-level encoder for global, high-level features, a pixel-level encoder for dense spatial features, and a training strategy combining SFT + GRPO-based reinforcement (Bai et al., 29 Sep 2025). The disease-level encoder is written as $e = f_{\mathrm{disease}}(I)$, where $I$ is the input image and $e$ is a global embedding. The pixel-level encoder is written as $$p \in \mathbb{R}^{H \times W \times C} = f_{\mathrm{pixel}}(I)$$, projected into the language-model space as $\hat p = \phi_{\mathrm{pixel}}(p)$.

This dual pathway is central to the model’s stated rationale. The disease-level pathway captures clinically salient global patterns for diagnosis/classification, while the pixel-level pathway retains dense spatial detail for fine-grained grounding and localization. The two streams are connected to a LLM through learnable multimodal connectors, with the disease-level connector written as $\hat e = \phi_{\mathrm{disease}}(e)$ and the pixel-level connector as $\hat p = \phi_{\mathrm{pixel}}(p)$ (Bai et al., 29 Sep 2025).

The language backbone is an LLM with LoRA-based adapters used for parameter-efficient fine-tuning, and the model size is stated to be approximately 8B parameters. Unified exemplar prompting conditions the LLM on both $\hat e$ and $\hat p$, along with an instruction or question $T$. The structured output format uses forms such as > ... followed by <answer> ... </answer>, so that reasoning tokens and grounded outputs are explicitly separated (Bai et al., 29 Sep 2025).

Several architectural details are deliberately left underspecified in the paper. The exact disease-level backbone architecture is not specified; the exact pixel-level backbone is not fully specified in the text; figure references mention “Medical SAM” and “SAM Connector”; and the input resolution, tokenizer scheme, and detailed connector parameterization are not provided. This leaves EVLF-FM defined more clearly at the functional level than at the low-level implementation level (Bai et al., 29 Sep 2025).

3. Training data, task coverage, and optimization strategy

The development corpus comprises 1,326,315 samples across three tasks. The diagnosis/classification component uses 554,391 samples across 9 modalities, sourced from MedMNIST—including PathMNIST, ChestMNIST, DermaMNIST, OCTMNIST, PneumoniaMNIST, RetinaMNIST, BreastMNIST, BloodMNIST, TissueMNIST, OrganAMNIST, OrganCMNIST, OrganSMNIST—together with BRSET, AIROGS, and VinDr CXR. The visual-grounding component uses 415,378 samples across 11 modalities, including BiomedParse and VinDr. The VQA component uses 208,230 samples from Path-VQA, VQA-RAD, SLAKE, and PMC-VQA (Bai et al., 29 Sep 2025).

The paper places this corpus across six clinical specialties—dermatology, hepatology, ophthalmology, pathology, pulmonology, and radiology—and up to eleven imaging modalities, including CT, MRI, X-ray, ultrasound, PET, histopathology images/slides, dermatoscopy, color fundus photography, optical coherence tomography, and endoscopy. External validation uses 8,884 independent test samples from 10 additional datasets across five imaging modalities, although the text explicitly lists nine datasets and notes that one dataset is unspecified (Bai et al., 29 Sep 2025).

Supervised fine-tuning is formulated through a unified negative log-likelihood objective,

$I$0

where the target response $I$1 is conditioned on the image, task instruction, and both projected visual streams (Bai et al., 29 Sep 2025).

The optimization schedule is explicitly two-stage. In Stage 1, the model performs supervised fine-tuning with LoRA: the core vision encoders and base LLM weights are frozen, while only the LoRA adapters and connector weights are updated. In Stage 2, the model performs reinforcement learning with GRPO (Group Relative Policy Optimization). The RL procedure samples groups of candidate outputs, computes group-average reward and normalized advantages, and applies task-specific verifiable rewards. For disease classification, the reward is $I$2 if the output contains the correct diagnosis in the expected format; otherwise $I$3. For pathology detection/localization, the reward is IoU(predicted region, ground truth), with reward zeroed when IoU is below a low threshold. For open-ended QA, the paper reports token-level F1 evaluation but does not apply RL directly because of the difficulty of automatic grading (Bai et al., 29 Sep 2025).

4. Empirical performance

EVLF-FM reports strong internal and external performance across its three task families. In internal validation for disease diagnostics, it achieves mean Accuracy = 0.858 and mean F1 = 0.797, outperforming listed generalist VLMs and specialist deep-learning baselines. In medical visual grounding, the average metrics across nine modalities are [email protected] = 0.949, [email protected] = 0.907, [email protected] = 0.837, and mIOU = 0.743. In external validation, it achieves top F1 and Accuracy on HRCTCov19 (0.954/0.949), RetinalOCT-C8 (0.481/0.338), and Glaucoma_fundus (0.740/0.732), while remaining competitive against larger models such as MedGemma 27B and MedDr 32B (Bai et al., 29 Sep 2025).

Evaluation setting	EVLF-FM result	Comparative note
Internal disease diagnostics	Mean Accuracy 0.858; mean F1 0.797	Outperforms listed generalist VLMs and specialist DL baselines
Internal visual grounding	[email protected] 0.837; mIOU 0.743	InternVL 8B baseline: [email protected] 0.755; mIOU 0.656
External Med-GRIT-270K grounding	Zero-shot [email protected] 36.17%; few-shot 54.08%	Few-shot surpasses BIRD’s 53.92% trained with the full dataset

Task-level detail reinforces this aggregate picture. For medical VQA, closed accuracy is reported as 89.6% on VQA-RAD, 91.5% on SLAKE, and 92.9% on Path-VQA. Open recall is 81.7% on VQA-RAD, 88.3% on SLAKE, and 23.7% on Path-VQA, with the paper noting that Path-VQA free-text grading is difficult across models (Bai et al., 29 Sep 2025).

The few-shot and data-efficiency results are also notable. For RetinalOCT, external performance improves from Accuracy/F1 0.481/0.338 in zero-shot mode to 0.697/0.604 with 64-shot conditioning. In the RL-data ablation, using 20% of the RL data yields Accuracy/F1 0.841/0.764, while the full setting reaches 0.858/0.797. The paper describes these differences as modest, indicating robustness to fine-tuning set size (Bai et al., 29 Sep 2025).

5. Explainability, grounding, and clinical role

The defining claim of EVLF-FM is not only multimodal breadth but also transparent reasoning paired with visual evidence. The model is trained to emit intermediate justifications that resemble clinical reasoning and to associate those justifications with grounded spatial outputs, typically bounding boxes. The paper explicitly contrasts this with black-box multimodal systems and argues that EVLF-FM aligns outputs with visual evidence through a combination of dense pixel-level encoding and IoU-based reward-guided fine-tuning (Bai et al., 29 Sep 2025).

This design has direct implications for faithfulness. Rather than relying on post hoc attention visualization as the core interpretability mechanism, EVLF-FM uses explicit localization outputs and evaluates them with [email protected]/0.3/0.5 and mIOU. The paper presents the model as producing reasoning and localization in a coupled form, so that a diagnostic answer can be inspected together with the region that allegedly supports it. A plausible implication is that the model’s interpretability claim rests less on narrative fluency than on the measurable agreement between textual reasoning targets and visual-grounding targets.

The reported clinical scope is broad. Demonstrated use cases span ophthalmology for diabetic-retinopathy grading and glaucoma detection, radiology for CT, MRI, and X-ray disease classification and localization, dermatology and endoscopy for lesion localization, and pathology for tissue classification and VQA. The paper also emphasizes human oversight, the risk of over-reliance on AI, the persistence of dataset bias, and the need for broader representative curation. It further states that local fine-tuning for specific settings could preserve privacy, although deployment metrics such as inference latency, memory footprint, or cost-effectiveness are not reported (Bai et al., 29 Sep 2025).

The limitations are correspondingly explicit. VQA performance lags diagnosis, especially for open-ended responses; generalizability and bias remain concerns; exact architectural and reproducibility details are incomplete; and code, model weights, licenses, seeds, and detailed hyperparameters are not provided in the paper. Statistical significance tests for the reported improvements are also not reported (Bai et al., 29 Sep 2025).

6. Disambiguation with other EVLF-FM usages

In the diffusion-based dataset-distillation literature, EVLF-FM does not denote a medical foundation model. The paper “EVLF: Early Vision-Language Fusion for Generative Dataset Distillation” states that it does not define a component explicitly named “EVLF-FM”; the appropriate interpretation of “FM” is the EVLF Fusion Module, a lightweight cross-attention module inserted once, before noise is added and before the denoiser is invoked, at the encoder-to-denoiser transition (Cai et al., 8 Mar 2026). That module aligns textual and visual embeddings so that local textures and global semantic directions co-evolve across the denoising process, and it is described as plug-and-play across UNet (LDM) and DiT backbones.

A separate cardiology-oriented interpretation uses the same string for echocardiographic LV flow field mapping. In that usage, EVLF-FM is realized by 3D-iVFM, a Doppler-based, physics-constrained reconstruction of the left-ventricular blood flow field that generalizes iVFM to full-volume, three-component estimation from a clinical triplane acquisition (Vixège et al., 2021). The formulation enforces mass conservation and free-slip boundary conditions, solves a least-squares problem with Lagrange multipliers, and was validated in silico against patient-specific CFD and tested in vivo on volunteers.

Another echocardiographic interpretation equates EVLF-FM with TEBVF(t), the total ejected blood volume fraction, introduced as a time-resolved generalization of conventional EF derived from volume changes of the left ventricular myocardial segments (1803.02825). In that formulation, EF is recovered at end-systole, while the generalized quantity integrates segmental contributions over time using longitudinal, circumferential, and radial strains together with segmental displacement.

These usages are conceptually unrelated. This suggests that EVLF-FM should not be treated as a stable cross-domain term; it requires expansion from local context before technical interpretation.