Andrea M. Bejar

Hongyi Pan, Gorkem Durak, Halil Ertugrul Aktas et al.

Publicly available full-field digital mammography (FFDM) datasets remain limited in size, clinical labels, and vendor diversity, which hinders the training of robust models. We present LUMINA, a curated, multi-vendor FFDM dataset that explicitly encodes acquisition energy and vendor metadata to expose clinically relevant appearance shifts that current benchmarks overlook. This innovative resource comprises 1824 images from 468 patients (960 benign, 864 malignant) with pathology-confirmed outcomes, BI-RADS assessments, and breast-density annotations. LUMINA spans six acquisition systems and both high- and low-energy styles, exposing vendor- and energy-driven appearance shifts. To reduce cross-vendor/energy drift while preserving lesion morphology, we introduce a foreground-only, pixel-space alignment (''energy harmonization'') that aligns each image to a low-energy reference style, leaving the zero-valued background unchanged. By benchmarking modern CNN and transformer baselines on three clinically meaningful tasks -- diagnosis (benign vs. malignant), BI-RADS risk grouping, and density -- we unify single-vs-two-view evaluation and show that two-view models consistently outperform single-view; in our benchmark, EfficientNet-B0 attains AUC 93.54% for diagnosis, and Swin-T yields the best macro-AUC 89.43% for density. Harmonization improves AUC/ACC across backbones and yields more focal Grad-CAM localization around suspicious regions. Being a richly annotated resource, LUMINA thus provides (a) a vendor-diverse, energy-labeled benchmark and (b) a model-agnostic harmonization protocol that together catalyze reliable, deployable mammography AI.

Zheyuan Zhang, Linkai Peng, Wanying Dou et al.

Clinical magnetic-resonance (MR) protocols generate many T1 and T2 sequences whose appearance differs more than the acquisition sites that produce them. Existing domain-generalization benchmarks focus almost on cross-center shifts and overlook this dominant source of variability. Pancreas segmentation remains a major challenge in abdominal imaging: the gland is small, irregularly, surrounded by organs and fat, and often suffers from low T1 contrast. State-of-the-art deep networks that already achieve >90% Dice on the liver or kidneys still miss 20-30% of the pancreas. The organ is also systematically under-represented in public cross-domain benchmarks, despite its clinical importance in early cancer detection, surgery, and diabetes research. To close this gap, we present PancreasDG, a large-scale multi-center 3D MRI pancreas segmentation dataset for investigating domain generalization in medical imaging. The dataset comprises 563 MRI scans from six institutions, spanning both venous phase and out-of-phase sequences, enabling study of both cross-center and cross-sequence variations with pixel-accurate pancreas masks created by a double-blind, two-pass protocol. Through comprehensive analysis, we reveal three insights: (i) limited sampling introduces significant variance that may be mistaken for distribution shifts, (ii) cross-center performance correlates with source domain performance for identical sequences, and (iii) cross-sequence shifts require specialized solutions. We also propose a semi-supervised approach that leverages anatomical invariances, significantly outperforming state-of-the-art domain generalization techniques with 61.63% Dice score improvements and 87.00% on two test centers for cross-sequence segmentation. PancreasDG sets a new benchmark for domain generalization in medical imaging. Dataset, code, and models will be available at https://pancreasdg.netlify.app.