arXiv subjects

Cosmin I. Bercea

Publications and source records attributed to Cosmin I. Bercea.

At least 19 recordsLinked to original sources

arXiv2026

RadPRISM: Schema-stratified radiology-report supervision for concept-disentangled image representations and visual grounding

Vision-language pretraining learns rich medical image representations from radiology reports, but previous model variants commonly operate within a single shared embedding space, so concept-level structure and interpretability must be recovered post hoc, limiting model transparency and, hence, clinical utility. We introduce RadPRISM, which makes a clinician-defined radiology schema a designated stratification axis: an on-premise large language model extracts per-concept text spans from free-text reports, and each clinical concept is aligned in its own dedicated visual subspace, turning concept stratification into direct, top-level alignment supervision. Instantiated on chest radiographs with a 19-concept schema over $203{,}602$ examinations from an internal multi-year archive, RadPRISM improved internal dataset zero-shot classification from $0.717$ (95% CI, $0.710-0.723$) to $0.868$ (95% CI, $0.863-0.872$) macro AUROC over a matched global-alignment baseline, performed on par with the purpose-built CARZero reference in external zero-shot classification while substantially outperforming it (up to 4.3-fold) in pointing-game visual grounding. In addition, a radiologist reader study demonstrated concept-stratified retrieval ability ($0.78$ macro retrieval correctness rate within rank 3), surfacing disentangled descriptive findings that report-level retrieval and fixed-label vocabularies cannot express. RadPRISM yields discriminative, spatially faithful, natively concept-stratified representations shaped by and transparently inspectable by clinicians.

cs.CV↗

arXiv2026

Dynamic Decision Learning: Test-Time Evolution for Abnormality Grounding in Rare Diseases

Clinical abnormality grounding for rare diseases is often hindered by data scarcity, making supervised fine-tuning impractical and single-pass inference highly unstable. We propose Dynamic Decision Learning (DDL), a framework that enables frozen large vision-language models (LVLMs) to refine their decisions across both language and visual spaces by optimizing instructions and consolidating predictions under visual perturbations. This process improves localization quality and produces a consensus-based reliability score that quantifies model confidence. Results on brain imaging benchmarks, including a rare-disease dataset with 281 pathology types across models ranging from 3B to 72B parameters, show that DDL improves mAP@75 by up to 105% on rare-disease cases and outperforms adaptation baselines and supervised fine-tuning. Furthermore, DDL demonstrates stronger calibration between reliability scores and localization accuracy under severe distribution shifts and increasing task difficulty. Code is available at: https://lijunrio.github.io/DDL/

cs.CL↗

arXiv2026

Measuring and Aligning Abstraction in Vision-Language Models with Medical Taxonomies

Vision-Language Models show strong zero-shot performance for chest X-ray classification, but standard flat metrics fail to distinguish between clinically minor and severe errors. This work investigates how to quantify and mitigate abstraction errors by leveraging medical taxonomies. We benchmark several state-of-the-art VLMs using hierarchical metrics and introduce Catastrophic Abstraction Errors to capture cross-branch mistakes. Our results reveal substantial misalignment of VLMs with clinical taxonomies despite high flat performance. To address this, we propose risk-constrained thresholding and taxonomy-aware fine-tuning with radial embeddings, which reduce severe abstraction errors to below 2 per cent while maintaining competitive performance. These findings highlight the importance of hierarchical evaluation and representation-level alignment for safer and more clinically meaningful deployment of VLMs.

cs.AI↗

arXiv2026

Does DINOv3 Set a New Medical Vision Standard? Benchmarking 2D and 3D Classification, Segmentation, and Registration

The advent of large-scale vision foundation models, pre-trained on diverse natural images, has marked a paradigm shift in computer vision. However, how the frontier vision foundation models' efficacies transfer to specialised domains such as medical imaging remains an open question. This report investigates whether DINOv3, a state-of-the-art self-supervised vision transformer (ViT) pre-trained on natural images, can directly serve as a powerful, unified encoder for medical vision tasks without domain-specific fine-tuning. To answer this, we benchmark DINOv3 across common medical vision tasks, including 2D and 3D classification, segmentation, and registration on a wide range of medical imaging modalities. We systematically analyse its scalability by varying model sizes and input image resolutions. Our findings reveal that DINOv3 shows impressive performance and establishes a formidable new baseline. Remarkably, it can even outperform medical-specific foundation models like BiomedCLIP and CT-Net on several tasks, despite being trained solely on natural images. However, we identify clear limitations: The model's features degrade in scenarios requiring deep domain specialisation, such as in whole-slide images (WSIs), electron microscopy (EM), and positron emission tomography (PET). Furthermore, we observe that DINOv3 does not consistently follow the scaling law in the medical domain. Its performance does not reliably increase with larger models or finer feature resolutions, showing diverse scaling behaviours across tasks. Overall, our work establishes DINOv3 as a strong baseline, whose powerful visual features can serve as a robust prior for multiple medical tasks. This opens promising future directions, such as leveraging its features to enforce multiview consistency in 3D reconstruction.

cs.CV↗

arXiv2025

TomoGraphView: 3D Medical Image Classification with Omnidirectional Slice Representations and Graph Neural Networks

The sharp rise in medical tomography examinations has created a demand for automated systems that can reliably extract informative features for downstream tasks such as tumor characterization. Although 3D volumes contain richer information than individual slices, effective 3D classification remains difficult: volumetric data encode complex spatial dependencies, and the scarcity of large-scale 3D datasets has constrained progress toward 3D foundation models. As a result, many recent approaches rely on 2D vision foundation models trained on natural images, repurposing them as feature extractors for medical scans with surprisingly strong performance. Despite their practical success, current methods that apply 2D foundation models to 3D scans via slice-based decomposition remain fundamentally limited. Standard slicing along axial, sagittal, and coronal planes often fails to capture the true spatial extent of a structure when its orientation does not align with these canonical views. More critically, most approaches aggregate slice features independently, ignoring the underlying 3D geometry and losing spatial coherence across slices. To overcome these limitations, we propose TomoGraphView, a novel framework that integrates omnidirectional volume slicing with spherical graph-based feature aggregation. Instead of restricting the model to axial, sagittal, or coronal planes, our method samples both canonical and non-canonical cross-sections generated from uniformly distributed points on a sphere enclosing the volume. We publicly share our accessible code base at http://github.com/compai-lab/2025-MedIA-kiechle and provide a user-friendly library for omnidirectional volume slicing at https://pypi.org/project/OmniSlicer.

eess.IV↗

arXiv2025

Denoising Diffusion Models for Anomaly Localization in Medical Images

This review explores anomaly localization in medical images using denoising diffusion models. After providing a brief methodological background of these models, including their application to image reconstruction and their conditioning using guidance mechanisms, we provide an overview of available datasets and evaluation metrics suitable for their application to anomaly localization in medical images. In this context, we discuss supervision schemes ranging from fully supervised segmentation to semi-supervised, weakly supervised, self-supervised, and unsupervised methods, and provide insights into the effectiveness and limitations of these approaches. Furthermore, we highlight open challenges in anomaly localization, including detection bias, domain shift, computational cost, and model interpretability. Our goal is to provide an overview of the current state of the art in the field, outline research gaps, and highlight the potential of diffusion models for robust anomaly localization in medical images.

eess.IV↗

arXiv2025

Learning to reason about rare diseases through retrieval-augmented agents

Rare diseases represent the long tail of medical imaging, where AI models often fail due to the scarcity of representative training data. In clinical workflows, radiologists frequently consult case reports and literature when confronted with unfamiliar findings. Following this line of reasoning, we introduce RADAR, Retrieval Augmented Diagnostic Reasoning Agents, an agentic system for rare disease detection in brain MRI. Our approach uses AI agents with access to external medical knowledge by embedding both case reports and literature using sentence transformers and indexing them with FAISS to enable efficient similarity search. The agent retrieves clinically relevant evidence to guide diagnostic decision making on unseen diseases, without the need of additional training. Designed as a model-agnostic reasoning module, RADAR can be seamlessly integrated with diverse large language models, consistently improving their rare pathology recognition and interpretability. On the NOVA dataset comprising 280 distinct rare diseases, RADAR achieves up to a 10.2% performance gain, with the strongest improvements observed for open source models such as DeepSeek. Beyond accuracy, the retrieved examples provide interpretable, literature grounded explanations, highlighting retrieval-augmented reasoning as a powerful paradigm for low-prevalence conditions in medical imaging.

cs.CL↗

arXiv2025

Influence of Classification Task and Distribution Shift Type on OOD Detection in Fetal Ultrasound

Reliable out-of-distribution (OOD) detection is important for safe deployment of deep learning models in fetal ultrasound amidst heterogeneous image characteristics and clinical settings. OOD detection relies on estimating a classification model's uncertainty, which should increase for OOD samples. While existing research has largely focused on uncertainty quantification methods, this work investigates the impact of the classification task itself. Through experiments with eight uncertainty quantification methods across four classification tasks, we demonstrate that OOD detection performance significantly varies with the task, and that the best task depends on the defined ID-OOD criteria; specifically, whether the OOD sample is due to: i) an image characteristic shift or ii) an anatomical feature shift. Furthermore, we reveal that superior OOD detection does not guarantee optimal abstained prediction, underscoring the necessity to align task selection and uncertainty strategies with the specific downstream application in medical image analysis.

cs.CV↗

arXiv2025

Knowledge to Sight: Reasoning over Visual Attributes via Knowledge Decomposition for Abnormality Grounding

In this work, we address the problem of grounding abnormalities in medical images, where the goal is to localize clinical findings based on textual descriptions. While generalist Vision-Language Models (VLMs) excel in natural grounding tasks, they often struggle in the medical domain due to rare, compositional, and domain-specific terms that are poorly aligned with visual patterns. Specialized medical VLMs address this challenge via large-scale domain pretraining, but at the cost of substantial annotation and computational resources. To overcome these limitations, we propose \textbf{Knowledge to Sight (K2Sight)}, a framework that introduces structured semantic supervision by decomposing clinical concepts into interpretable visual attributes, such as shape, density, and anatomical location. These attributes are distilled from domain ontologies and encoded into concise instruction-style prompts, which guide region-text alignment during training. Unlike conventional report-level supervision, our approach explicitly bridges domain knowledge and spatial structure, enabling data-efficient training of compact models. We train compact models with 0.23B and 2B parameters using only 1.5\% of the data required by state-of-the-art medical VLMs. Despite their small size and limited training data, these models achieve performance on par with or better than 7B+ medical VLMs, with up to 9.82\% improvement in $mAP_{50}$. Code and models: \href{https://lijunrio.github.io/K2Sight/}{\textcolor{SOTAPink}{https://lijunrio.github.io/K2Sight/}}.

cs.CV↗

arXiv2025

NOVA: A Benchmark for Anomaly Localization and Clinical Reasoning in Brain MRI

In many real-world applications, deployed models encounter inputs that differ from the data seen during training. Out-of-distribution detection identifies whether an input stems from an unseen distribution, while open-world recognition flags such inputs to ensure the system remains robust as ever-emerging, previously $unknown$ categories appear and must be addressed without retraining. Foundation and vision-language models are pre-trained on large and diverse datasets with the expectation of broad generalization across domains, including medical imaging. However, benchmarking these models on test sets with only a few common outlier types silently collapses the evaluation back to a closed-set problem, masking failures on rare or truly novel conditions encountered in clinical use. We therefore present $NOVA$, a challenging, real-life $evaluation-only$ benchmark of $\sim$900 brain MRI scans that span 281 rare pathologies and heterogeneous acquisition protocols. Each case includes rich clinical narratives and double-blinded expert bounding-box annotations. Together, these enable joint assessment of anomaly localisation, visual captioning, and diagnostic reasoning. Because NOVA is never used for training, it serves as an $extreme$ stress-test of out-of-distribution generalisation: models must bridge a distribution gap both in sample appearance and in semantic space. Baseline results with leading vision-language models (GPT-4o, Gemini 2.0 Flash, and Qwen2.5-VL-72B) reveal substantial performance drops across all tasks, establishing NOVA as a rigorous testbed for advancing models that can detect, localize, and reason about truly unknown anomalies.

eess.IV↗

arXiv2025

Semantic Alignment of Unimodal Medical Text and Vision Representations

General-purpose AI models, particularly those designed for text and vision, demonstrate impressive versatility across a wide range of deep-learning tasks. However, they often underperform in specialised domains like medical imaging, where domain-specific solutions or alternative knowledge transfer approaches are typically required. Recent studies have noted that general-purpose models can exhibit similar latent spaces when processing semantically related data, although this alignment does not occur naturally. Building on this insight, it has been shown that applying a simple transformation - at most affine - estimated from a subset of semantically corresponding samples, known as anchors, enables model stitching across diverse training paradigms, architectures, and modalities. In this paper, we explore how semantic alignment - estimating transformations between anchors - can bridge general-purpose AI with specialised medical knowledge. Using multiple public chest X-ray datasets, we demonstrate that model stitching across model architectures allows general models to integrate domain-specific knowledge without additional training, leading to improved performance on medical tasks. Furthermore, we introduce a novel zero-shot classification approach for unimodal vision encoders that leverages semantic alignment across modalities. Our results show that our method not only outperforms general multimodal models but also approaches the performance levels of fully trained, medical-specific multimodal solutions

cs.CV↗

arXiv2025

Enhancing Abnormality Grounding for Vision Language Models with Knowledge Descriptions

Visual Language Models (VLMs) have demonstrated impressive capabilities in visual grounding tasks. However, their effectiveness in the medical domain, particularly for abnormality detection and localization within medical images, remains underexplored. A major challenge is the complex and abstract nature of medical terminology, which makes it difficult to directly associate pathological anomaly terms with their corresponding visual features. In this work, we introduce a novel approach to enhance VLM performance in medical abnormality detection and localization by leveraging decomposed medical knowledge. Instead of directly prompting models to recognize specific abnormalities, we focus on breaking down medical concepts into fundamental attributes and common visual patterns. This strategy promotes a stronger alignment between textual descriptions and visual features, improving both the recognition and localization of abnormalities in medical images.We evaluate our method on the 0.23B Florence-2 base model and demonstrate that it achieves comparable performance in abnormality grounding to significantly larger 7B LLaVA-based medical VLMs, despite being trained on only 1.5% of the data used for such models. Experimental results also demonstrate the effectiveness of our approach in both known and previously unseen abnormalities, suggesting its strong generalization capabilities.

cs.CV↗

arXiv2024

Selective Test-Time Adaptation for Unsupervised Anomaly Detection using Neural Implicit Representations

Deep learning models in medical imaging often encounter challenges when adapting to new clinical settings unseen during training. Test-time adaptation offers a promising approach to optimize models for these unseen domains, yet its application in anomaly detection (AD) remains largely unexplored. AD aims to efficiently identify deviations from normative distributions; however, full adaptation, including pathological shifts, may inadvertently learn the anomalies it intends to detect. We introduce a novel concept of selective test-time adaptation that utilizes the inherent characteristics of deep pre-trained features to adapt selectively in a zero-shot manner to any test image from an unseen domain. This approach employs a model-agnostic, lightweight multi-layer perceptron for neural implicit representations, enabling the adaptation of outputs from any reconstruction-based AD method without altering the source-trained model. Rigorous validation in brain AD demonstrated that our strategy substantially enhances detection accuracy for multiple conditions and different target distributions. Specifically, our method improves the detection rates by up to 78% for enlarged ventricles and 24% for edemas.

cs.LG↗

arXiv2024

FMBench: Benchmarking Fairness in Multimodal Large Language Models on Medical Tasks

Advancements in Multimodal Large Language Models (MLLMs) have significantly improved medical task performance, such as Visual Question Answering (VQA) and Report Generation (RG). However, the fairness of these models across diverse demographic groups remains underexplored, despite its importance in healthcare. This oversight is partly due to the lack of demographic diversity in existing medical multimodal datasets, which complicates the evaluation of fairness. In response, we propose FMBench, the first benchmark designed to evaluate the fairness of MLLMs performance across diverse demographic attributes. FMBench has the following key features: 1: It includes four demographic attributes: race, ethnicity, language, and gender, across two tasks, VQA and RG, under zero-shot settings. 2: Our VQA task is free-form, enhancing real-world applicability and mitigating the biases associated with predefined choices. 3: We utilize both lexical metrics and LLM-based metrics, aligned with clinical evaluations, to assess models not only for linguistic accuracy but also from a clinical perspective. Furthermore, we introduce a new metric, Fairness-Aware Performance (FAP), to evaluate how fairly MLLMs perform across various demographic attributes. We thoroughly evaluate the performance and fairness of eight state-of-the-art open-source MLLMs, including both general and medical MLLMs, ranging from 7B to 26B parameters on the proposed benchmark. We aim for FMBench to assist the research community in refining model evaluation and driving future advancements in the field. All data and code will be released upon acceptance.

cs.CV↗

arXiv2024

Interpretable Representation Learning of Cardiac MRI via Attribute Regularization

Interpretability is essential in medical imaging to ensure that clinicians can comprehend and trust artificial intelligence models. Several approaches have been recently considered to encode attributes in the latent space to enhance its interpretability. Notably, attribute regularization aims to encode a set of attributes along the dimensions of a latent representation. However, this approach is based on Variational AutoEncoder and suffers from blurry reconstruction. In this paper, we propose an Attributed-regularized Soft Introspective Variational Autoencoder that combines attribute regularization of the latent space within the framework of an adversarially trained variational autoencoder. We demonstrate on short-axis cardiac Magnetic Resonance images of the UK Biobank the ability of the proposed method to address blurry reconstruction issues of variational autoencoder methods while preserving the latent space interpretability.

eess.IV↗

arXiv2024

Language Models Meet Anomaly Detection for Better Interpretability and Generalizability

This research explores the integration of language models and unsupervised anomaly detection in medical imaging, addressing two key questions: (1) Can language models enhance the interpretability of anomaly detection maps? and (2) Can anomaly maps improve the generalizability of language models in open-set anomaly detection tasks? To investigate these questions, we introduce a new dataset for multi-image visual question-answering on brain magnetic resonance images encompassing multiple conditions. We propose KQ-Former (Knowledge Querying Transformer), which is designed to optimally align visual and textual information in limited-sample contexts. Our model achieves a 60.81% accuracy on closed questions, covering disease classification and severity across 15 different classes. For open questions, KQ-Former demonstrates a 70% improvement over the baseline with a BLEU-4 score of 0.41, and achieves the highest entailment ratios (up to 71.9%) and lowest contradiction ratios (down to 10.0%) among various natural language inference models. Furthermore, integrating anomaly maps results in an 18% accuracy increase in detecting open-set anomalies, thereby enhancing the language model's generalizability to previously unseen medical conditions. The code and dataset are available at https://github.com/compai-lab/miccai-2024-junli?tab=readme-ov-file

cs.CV↗

arXiv2024

Diffusion Models for Unsupervised Anomaly Detection in Fetal Brain Ultrasound

Ultrasonography is an essential tool in mid-pregnancy for assessing fetal development, appreciated for its non-invasive and real-time imaging capabilities. Yet, the interpretation of ultrasound images is often complicated by acoustic shadows, speckle noise, and other artifacts that obscure crucial diagnostic details. To address these challenges, our study presents a novel unsupervised anomaly detection framework specifically designed for fetal ultrasound imaging. This framework incorporates gestational age filtering, precise identification of fetal standard planes, and targeted segmentation of brain regions to enhance diagnostic accuracy. Furthermore, we introduce the use of denoising diffusion probabilistic models in this context, marking a significant innovation in detecting previously unrecognized anomalies. We rigorously evaluated the framework using various diffusion-based anomaly detection methods, noise types, and noise levels. Notably, AutoDDPM emerged as the most effective, achieving an area under the precision-recall curve of 79.8\% in detecting anomalies. This advancement holds promise for improving the tools available for nuanced and effective prenatal diagnostics.

eess.IV↗

arXiv2024

MedEdit: Counterfactual Diffusion-based Image Editing on Brain MRI

Denoising diffusion probabilistic models enable high-fidelity image synthesis and editing. In biomedicine, these models facilitate counterfactual image editing, producing pairs of images where one is edited to simulate hypothetical conditions. For example, they can model the progression of specific diseases, such as stroke lesions. However, current image editing techniques often fail to generate realistic biomedical counterfactuals, either by inadequately modeling indirect pathological effects like brain atrophy or by excessively altering the scan, which disrupts correspondence to the original images. Here, we propose MedEdit, a conditional diffusion model for medical image editing. MedEdit induces pathology in specific areas while balancing the modeling of disease effects and preserving the integrity of the original scan. We evaluated MedEdit on the Atlas v2.0 stroke dataset using Frechet Inception Distance and Dice scores, outperforming state-of-the-art diffusion-based methods such as Palette (by 45%) and SDEdit (by 61%). Additionally, clinical evaluations by a board-certified neuroradiologist confirmed that MedEdit generated realistic stroke scans indistinguishable from real ones. We believe this work will enable counterfactual image editing research to further advance the development of realistic and clinically useful imaging tools.

eess.IV↗