A flaw in how deep learning models train on brain scans turns out to be exploitable as a strength.
Researchers studying FastSurfer, a popular deep learning model for brain segmentation, found that retraining it produces meaningfully different results each time - a phenomenon driven by numerical uncertainty baked into the training process. That variability exceeds what you get from FreeSurfer 7.3.2, the older non-deep-learning tool it is often compared against, particularly in cortical regions. The team traced the instability to random seed perturbations and confirmed that seed-induced variability closely mirrors the broader numerical variability seen across training runs.
The catch - and the upside - is that this variability is not pure noise. By ensembling models trained with different random seeds, the researchers improved performance on a downstream brain age regression task. In other words, the same instability that could skew a clinical result can, if harnessed deliberately, make predictions more accurate. That reframe matters for any lab relying on deep learning segmentation as a stepping stone to clinical outputs.
Neuroimaging is one of the higher-stakes corners of applied AI, where a model variance that looks like a rounding error on a benchmark can translate into a misread biomarker. The finding joins a small but growing body of work pushing back on the assumption that newer deep learning pipelines are simply better than their classical predecessors - they are often faster, but reliability is a separate question.