CRG Score: A Distribution-Aware Clinical Metric for Radiology Report Generation

Evaluating long-context radiology report generation is challenging. NLG metrics fail to capture clinical correctness, while LLM-based metrics often lack generalizability. Clinical accuracy metrics are more relevant but are sensitive to class imbalance, frequently favoring trivial predictions. We propose the CRG Score, a distribution-aware and adaptable metric that evaluates only clinically relevant abnormalities explicitly described in reference reports. CRG supports both binary and structured labels (e.g., type, location) and can be paired with any LLM for feature extraction. By balancing penalties based on label distribution, it enables fairer, more robust evaluation and serves as a clinically aligned reward function.

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.

Leveraging Multi-Modal Information to Enhance Dataset Distillation

Dataset distillation aims to create a compact and highly representative synthetic dataset that preserves the knowledge of a larger real dataset. While existing methods primarily focus on optimizing visual representations, incorporating additional modalities and refining object-level information can significantly improve the quality of distilled datasets. In this work, we introduce two key enhancements to dataset distillation: caption-guided supervision and object-centric masking. To integrate textual information, we propose two strategies for leveraging caption features: the feature concatenation, where caption embeddings are fused with visual features at the classification stage, and caption matching, which introduces a caption-based alignment loss during training to ensure semantic coherence between real and synthetic data. Additionally, we apply segmentation masks to isolate target objects and remove background distractions, introducing two loss functions designed for object-centric learning: masked feature alignment loss and masked gradient matching loss. Comprehensive evaluations demonstrate that integrating caption-based guidance and object-centric masking enhances dataset distillation, leading to synthetic datasets that achieve superior performance on downstream tasks.

Video Dataset Condensation with Diffusion Models

In recent years, the rapid expansion of dataset sizes and the increasing complexity of deep learning models have significantly escalated the demand for computational resources, both for data storage and model training. Dataset distillation has emerged as a promising solution to address this challenge by generating a compact synthetic dataset that retains the essential information from a large real dataset. However, existing methods often suffer from limited performance and poor data quality, particularly in the video domain. In this paper, we focus on video dataset distillation by employing a video diffusion model to generate high-quality synthetic videos. To enhance representativeness, we introduce Video Spatio-Temporal U-Net (VST-UNet), a model designed to select a diverse and informative subset of videos that effectively captures the characteristics of the original dataset. To further optimize computational efficiency, we explore a training-free clustering algorithm, Temporal-Aware Cluster-based Distillation (TAC-DT), to select representative videos without requiring additional training overhead. We validate the effectiveness of our approach through extensive experiments on four benchmark datasets, demonstrating performance improvements of up to \(10.61\%\) over the state-of-the-art. Our method consistently outperforms existing approaches across all datasets, establishing a new benchmark for video dataset distillation.

Dataset Distillation with Probabilistic Latent Features

As deep learning models grow in complexity and the volume of training data increases, reducing storage and computational costs becomes increasingly important. Dataset distillation addresses this challenge by synthesizing a compact set of synthetic data that can effectively replace the original dataset in downstream classification tasks. While existing methods typically rely on mapping data from pixel space to the latent space of a generative model, we propose a novel stochastic approach that models the joint distribution of latent features. This allows our method to better capture spatial structures and produce diverse synthetic samples, which benefits model training. Specifically, we introduce a low-rank multivariate normal distribution parameterized by a lightweight network. This design maintains low computational complexity and is compatible with various matching networks used in dataset distillation. After distillation, synthetic images are generated by feeding the learned latent features into a pretrained generator. These synthetic images are then used to train classification models, and performance is evaluated on real test set. We validate our method on several benchmarks, including ImageNet subsets, CIFAR-10, and the MedMNIST histopathological dataset. Our approach achieves state-of-the-art cross architecture performance across a range of backbone architectures, demonstrating its generality and effectiveness.

SpinMeRound: Consistent Multi-View Identity Generation Using Diffusion Models

Despite recent progress in diffusion models, generating realistic head portraits from novel viewpoints remains a significant challenge. Most current approaches are constrained to limited angular ranges, predominantly focusing on frontal or near-frontal views. Moreover, although the recent emerging large-scale diffusion models have been proven robust in handling 3D scenes, they underperform on facial data, given their complex structure and the uncanny valley pitfalls. In this paper, we propose SpinMeRound, a diffusion-based approach designed to generate consistent and accurate head portraits from novel viewpoints. By leveraging a number of input views alongside an identity embedding, our method effectively synthesizes diverse viewpoints of a subject whilst robustly maintaining its unique identity features. Through experimentation, we showcase our model's generation capabilities in 360 head synthesis, while beating current state-of-the-art multiview diffusion models.

EchoFlow: A Foundation Model for Cardiac Ultrasound Image and Video Generation

Advances in deep learning have significantly enhanced medical image analysis, yet the availability of large-scale medical datasets remains constrained by patient privacy concerns. We present EchoFlow, a novel framework designed to generate high-quality, privacy-preserving synthetic echocardiogram images and videos. EchoFlow comprises four key components: an adversarial variational autoencoder for defining an efficient latent representation of cardiac ultrasound images, a latent image flow matching model for generating accurate latent echocardiogram images, a latent re-identification model to ensure privacy by filtering images anatomically, and a latent video flow matching model for animating latent images into realistic echocardiogram videos conditioned on ejection fraction. We rigorously evaluate our synthetic datasets on the clinically relevant task of ejection fraction regression and demonstrate, for the first time, that downstream models trained exclusively on EchoFlow-generated synthetic datasets achieve performance parity with models trained on real datasets. We release our models and synthetic datasets, enabling broader, privacy-compliant research in medical ultrasound imaging at https://huggingface.co/spaces/HReynaud/EchoFlow.

L-FUSION: Laplacian Fetal Ultrasound Segmentation & Uncertainty Estimation

Accurate analysis of prenatal ultrasound (US) is essential for early detection of developmental anomalies. However, operator dependency and technical limitations (e.g. intrinsic artefacts and effects, setting errors) can complicate image interpretation and the assessment of diagnostic uncertainty. We present L-FUSION (Laplacian Fetal US Segmentation with Integrated FoundatiON models), a framework that integrates uncertainty quantification through unsupervised, normative learning and large-scale foundation models for robust segmentation of fetal structures in normal and pathological scans. We propose to utilise the aleatoric logit distributions of Stochastic Segmentation Networks and Laplace approximations with fast Hessian estimations to estimate epistemic uncertainty only from the segmentation head. This enables us to achieve reliable abnormality quantification for instant diagnostic feedback. Combined with an integrated Dropout component, L-FUSION enables reliable differentiation of lesions from normal fetal anatomy with enhanced uncertainty maps and segmentation counterfactuals in US imaging. It improves epistemic and aleatoric uncertainty interpretation and removes the need for manual disease-labelling. Evaluations across multiple datasets show that L-FUSION achieves superior segmentation accuracy and consistent uncertainty quantification, supporting on-site decision-making and offering a scalable solution for advancing fetal ultrasound analysis in clinical settings.

Image Generation Diversity Issues and How to Tame Them

Generative methods now produce outputs nearly indistinguishable from real data but often fail to fully capture the data distribution. Unlike quality issues, diversity limitations in generative models are hard to detect visually, requiring specific metrics for assessment. In this paper, we draw attention to the current lack of diversity in generative models and the inability of common metrics to measure this. We achieve this by framing diversity as an image retrieval problem, where we measure how many real images can be retrieved using synthetic data as queries. This yields the Image Retrieval Score (IRS), an interpretable, hyperparameter-free metric that quantifies the diversity of a generative model's output. IRS requires only a subset of synthetic samples and provides a statistical measure of confidence. Our experiments indicate that current feature extractors commonly used in generative model assessment are inadequate for evaluating diversity effectively. Consequently, we perform an extensive search for the best feature extractors to assess diversity. Evaluation reveals that current diffusion models converge to limited subsets of the real distribution, with no current state-of-the-art models superpassing 77% of the diversity of the training data. To address this limitation, we introduce Diversity-Aware Diffusion Models (DiADM), a novel approach that improves diversity of unconditional diffusion models without loss of image quality. We do this by disentangling diversity from image quality by using a diversity aware module that uses pseudo-unconditional features as input. We provide a Python package offering unified feature extraction and metric computation to further facilitate the evaluation of generative models https://github.com/MischaD/beyondfid.

Uncovering Hidden Subspaces in Video Diffusion Models Using Re-Identification

Latent Video Diffusion Models can easily deceive casual observers and domain experts alike thanks to the produced image quality and temporal consistency. Beyond entertainment, this creates opportunities around safe data sharing of fully synthetic datasets, which are crucial in healthcare, as well as other domains relying on sensitive personal information. However, privacy concerns with this approach have not fully been addressed yet, and models trained on synthetic data for specific downstream tasks still perform worse than those trained on real data. This discrepancy may be partly due to the sampling space being a subspace of the training videos, effectively reducing the training data size for downstream models. Additionally, the reduced temporal consistency when generating long videos could be a contributing factor. In this paper, we first show that training privacy-preserving models in latent space is computationally more efficient and generalize better. Furthermore, to investigate downstream degradation factors, we propose to use a re-identification model, previously employed as a privacy preservation filter. We demonstrate that it is sufficient to train this model on the latent space of the video generator. Subsequently, we use these models to evaluate the subspace covered by synthetic video datasets and thus introduce a new way to measure the faithfulness of generative machine learning models. We focus on a specific application in healthcare echocardiography to illustrate the effectiveness of our novel methods. Our findings indicate that only up to 30.8% of the training videos are learned in latent video diffusion models, which could explain the lack of performance when training downstream tasks on synthetic data.