Uncertain but Useful: Leveraging CNN Variability into Data Augmentation

Deep learning (DL) is rapidly advancing neuroimaging by achieving state-of-the-art performance with reduced computation times. Yet the numerical stability of DL models -- particularly during training -- remains underexplored. While inference with DL is relatively stable, training introduces additional variability primarily through iterative stochastic optimization. We investigate this training-time variability using FastSurfer, a CNN-based whole-brain segmentation pipeline. Controlled perturbations are introduced via floating point perturbations and random seeds. We find that: (i) FastSurfer exhibits higher variability compared to that of a traditional neuroimaging pipeline, suggesting that DL inherits and is particularly susceptible to sources of instability present in its predecessors; (ii) ensembles generated with perturbations achieve performance similar to an unperturbed baseline; and (iii) variability effectively produces ensembles of numerical model families that can be repurposed for downstream applications. As a proof of concept, we demonstrate that numerical ensembles can be used as a data augmentation strategy for brain age regression. These findings position training-time variability not only as a reproducibility concern but also as a resource that can be harnessed to improve robustness and enable new applications in neuroimaging.

Numerical Uncertainty of Convolutional Neural Networks Inference for Structural Brain MRI Analysis

This paper investigates the numerical uncertainty of Convolutional Neural Networks (CNNs) inference for structural brain MRI analysis. It applies Random Rounding -- a stochastic arithmetic technique -- to CNN models employed in non-linear registration (SynthMorph) and whole-brain segmentation (FastSurfer), and compares the resulting numerical uncertainty to the one measured in a reference image-processing pipeline (FreeSurfer recon-all). Results obtained on 32 representative subjects show that CNN predictions are substantially more accurate numerically than traditional image-processing results (non-linear registration: 19 vs 13 significant bits on average; whole-brain segmentation: 0.99 vs 0.92 S{\o}rensen-Dice score on average), which suggests a better reproducibility of CNN results across execution environments.