When the full picture is missing, medical AI should say so — not guess confidently.
Researchers have published a framework called SIUM that rethinks how deep learning models handle incomplete MRI data during brain tumor segmentation. Most clinical workflows acquire multiple MRI modalities — T1, T2, FLAIR, and others — but in practice some scans are missing at inference time due to motion artifacts, scanner limitations, or time constraints. Existing models tend to fill that gap with deterministic outputs that look precise but quietly paper over the missing information. SIUM instead represents each prediction as a probability distribution: the mean carries the actual tumor-boundary estimate, and the variance encodes how much the model does not know because of missing data. The team also built in a hierarchical ordering constraint so that a model working with three out of four modalities is always less uncertain than one working with two, keeping the uncertainty scores internally consistent. Tests on the BraTS 2018 and 2020 benchmarks showed the approach outperformed baselines across a range of missing-modality scenarios.
The reliability gap in medical AI is not a niche concern. Regulators and clinicians increasingly want models that can distinguish confident outputs from educated guesses, and radiology — where scan availability varies by institution and urgency — is one of the hardest places to enforce complete data pipelines. A model that surfaces its own uncertainty gives a radiologist something to act on; one that hides it is a liability dressed as a tool.
Code and model weights are public on GitHub, which is the right call — but BraTS is a well-worn benchmark, and real-world MRI distributions are messier than any curated dataset.