As part of this paper, we also present the most extensive framework for the meta-evaluation of medical image similarity metrics to date, available at github.com/mazurowski-l...!

FRD improves on other metrics (FID, RadiologyFID, KID, etc.) in many applications, including: OOD detection, image-to-image translation/image generation evaluation, correlation with expert-perceived image quality, compute stability+speed, and sensitivity to adversarial attacks + image corruptions.

FRD quantifies images via hundreds of standardized, interpretable radiomic features, rather than learned embeddings, which we find brings several advantages due to better and more robustly characterizing anatomical features (particularly those related to downstream tasks).

FID has many limitations when applied to medical images (e.g., misalignment with downstream tasks). Introducing the Fréchet Radiomic Distance (FRD): a metric designed from the ground up for better characterizing medical images, led by me and Richard Osuala! Code/paper: github.com/RichardObi/f...

With a novel dynamic "short-long" memory, it outperforms SAM 2 and other models by >7% DSC on average, in segmenting various organs, bones, and muscles across modalities! Crucially, it exhibits better robustness to the problem of "over-propagation" of annotations through slices.

Introducing SLM-SAM 2, led by Yuwen Chen: a new video object segmentation method for medical imaging that speeds up annotation by accurately propagating labels from a single slice across the whole volume. [Code: github.com/mazurowski-l..., paper: www.arxiv.org/abs/2505.01854 ]. More info next:

I also found that the representation intrinsic dim. peaks consistently earlier in medical image models compared to natural image models, pointing to a difference in the abstractness of task-relevant features between these domains. (2/3)

Excited to share my NeurIPS SciForDL paper on how hidden repr. intrinsic dimension evolves through model depth! openreview.net/forum?id=trc...
A surprising finding: the peak intrinsic dim. is ~a constant fraction of the input data’s intrinsic dim., across diverse datasets + models! (1/3)