News & Updates

We are pleased to announce that MOSCARD — our framework for causal de-confounding in multimodal AI — has been accepted to MICCAI 2025. MOSCARD addresses a fundamental challenge in multimodal fusion: separating true predictive features from spurious correlations introduced by demographic variables and acquisition biases. The paper demonstrates significant improvements in model calibration and fairness across subgroups in cardiovascular opportunistic screening tasks. Congratulations to the entire development team!

We are proud to report the completion of MEFINDER Milestone 1 ahead of schedule. All six planned open-source tool deliverables have been released on GitHub with full documentation, unit tests, and reproducibility notebooks. The tools collectively form a complete preprocessing and feature extraction pipeline for both breast and prostate cancer multimodal analysis. We thank all consortium members for their contributions to this milestone.

MamoCLIP introduces a novel application of CLIP-style contrastive learning to full-field digital mammography, trained in a federated setting across institutions in the EMBED v2 dataset. The model produces transferable representations that outperform supervised baselines on breast density classification and lesion detection tasks, without requiring centralization of patient data.

In a major step for open science in breast imaging AI, EMBED v2 is now publicly available through TCIA. The dataset includes digital mammography (FFDM and DBT), breast MRI, and linked clinical records for over 260,000 patients, making it the largest publicly accessible institutional breast imaging dataset in the world. Access requires a data use agreement and IRB review.

The RSNA Annual Meeting was a productive showcase for MEFINDER research. Our team presented five posters and one oral presentation covering MamoCLIP results, MQUAL validation on the VA dataset, F-SYN stain normalization benchmarks, ProstateNet segmentation performance, and an overview of the MEFINDER multimodal fusion framework. We thank all attendees who engaged with our work.

APIC (AI-based Pathology Image Classifier) represents a significant step toward affordable prognostics in prostate cancer. The paper demonstrates that APIC can predict treatment benefit from standard H&E slides with performance comparable to the Decipher molecular assay, at a fraction of the cost. Validation on CHAARTED and STAMPEDE clinical trial datasets provides strong evidence of generalizability.