Annotated medical imaging data remain scarce, and labels are often weak and noisy: in mammography, an examination comprises several high-resolution views, yet the diagnostic outcome is recorded only at the breast level. Such problems are naturally cast as Multiple Instance Learning (MIL), where the model must infer instance-level structure from bag-level labels alone. Although contemporary foundation encoders supply strong general-purpose embeddings, augmenting MIL data in this representation space remains an open problem as established techniques act on one instance at a time and ignore the statistical dependencies binding a bag together. We address this with SetFlow, a generative model that learns the distribution of complete MIL bags directly in a frozen encoder’s embedding space. SetFlow couples flow-matching training with a Set Transformer-inspired backbone, making it invariant to instance ordering while modeling intra-bag relationships. Generation is conditioned jointly on class label and per-instance scale, yielding coherent, semantically faithful bags rather than isolated vectors. Evaluating on two large public mammography datasets and two encoders, we assess distributional fidelity, nearest-neighbor behavior, and downstream augmentation utility. We show that generated bags reproduce real-data statistics and improve classification in certain configuration, with performance gains varying on the amount of synthetic data. An architecture ablation confirms each design choice contributes to performance.

Computed tomography (CT) is widely used for evaluating renal anatomy and pathology, with multi-phase contrastenhanced imaging providing critical diagnostic information. However, acquiring multiple contrast phases requires repeated scans, in turn increasing radiation exposure. In addition, the development of AI methods for medical imaging is limited by the scarcity of datasets, often due to privacy and data sensitivity concerns. To address both of these limitations, we propose PhaseDiff, which learns the relationship between non-contrast and contrast-enhanced renal CT scans to synthesize arterial, portal venous, and delayed phases directly from non-contrast inputs. The framework is based on a conditional latent diffusion model that operates in a learned latent space, where a variational autoencoder encodes images into compact latent representations. Diffusion is applied only to the target contrast-enhanced latent, while the non-contrast latent is incorporated via channel-wise concatenation to provide spatial guidance. In addition, the desired contrast phase is introduced through cross-attention, enabling phase-controlled image synthesis. Trained on paired multi-phase CT data, PhaseDiff generates anatomically and phase-consistent contrast-enhanced images, highlighting the potential to reduce radiation exposure in renal CT workflows while also creating realistic synthetic datasets.

Data scarcity and weak supervision continue to limit the performance of machine learning models in many real-world applications, such as mammography, where Multiple Instance Learning (MIL) often offers the best formulation. While recent foundation models provide strong semantic representations out of the box, effective augmentation of such representations of MIL data remains limited, as existing methods operate at the instance level and fail to capture intra-bag dependencies. In this work, we introduce SetFlow, a generative architecture that models entire MIL bags (i.e., sets) directly in the representation space. Our approach leverages the flow matching paradigm combined with a Set Transformer-inspired design, enabling it to handle permutation-invariant inputs while capturing interactions between instances within each bag. The model is conditioned on both class labels and input scale, allowing it to generate coherent and semantically consistent sets of representations. We evaluate SetFlow on a large-scale mammography benchmark using a stateof-the-art MIL-PF classification pipeline. The generated samples are shown to closely match the original data distribution and even improve downstream performance when used for augmentation. Furthermore, training on synthetic data alone shows competitive results, demonstrating the effectiveness of representation-space generative modeling for data-scarce and privacy-sensitive tasks.

Mammography is the leading methodology used to diagnose breast cancer. Effective, cheap and reliable, the mammography can be used to screen large populations, if the imagery produced can be analysed efficiently. State-of-the-art generative artificial intelligence approaches can be used to create tools able to aid in this task. Here we present a study focused on the emerging research topic of the application of generative diffusion models to the task of anomaly detection and we apply if for detecting anomalies on mammograms. Diffusion models exhibit promising results in making pixel-level predictions with image level annotations, but no such application has been published so far regarding mammography. We have, therefore, developed a novel approach utilizing U-net backbone that is able to generate mammograms with Fréchet Inception Distance (FID) of 14.62. We showed its ability to perform anomaly detection with Intersection over Union (IoU) of 0.195 which demonstrates the viability of our approach for early-stage research.

Svake godine u saobraćajnim nesrećama na putevima širom svijeta pogine 1,35 i bude povrijeđeno 20-50 miliona ljudi, što znači da svakog dana u prosjeku skoro 3.700 ljudi izgubi život u saobraćaju. Više od polovine poginulih su pješaci, motoristi ili biciklisti. U ovom radu će biti analizirani faktori koji direktno utiču na povećanje vjerovatnoće pojave sudara motornih vozila, a zatim predstavljena metodologija za predviđanje rizika kao i klasifikaciju nivoa ozbiljnosti sudara korištenjem klasifikacionih modela mašinskog učenja. Predloženi model objedinjuje podatke o sudarima, ulicama Njujorka, protoku saobraćaja na pojedinim dionicama i podatke o vremenu i može se koristiti za identifikaciju gdje i kada je rizik od nesreće značajno veći od prosjeka kako bi se preduzele radnje za smanjenje tog rizika. Rezultati modela za predikciju sudara dostižu tačnost od 70%, dok model za klasifikaciju ozbiljnosti sudara postiže makro-prosječni F1-skor od 0,56.