Conservative & Aggressive NaNs Accelerate U-Nets for Neuroimaging

Deep learning models for neuroimaging increasingly rely on large architectures, making efficiency a persistent concern despite advances in hardware. Through an analysis of numerical uncertainty of convolutional neural networks (CNNs), we observe that many operations are applied to values dominated by numerical noise and have negligible influence on model outputs. In some models, up to two-thirds of convolution operations appear redundant. We introduce Conservative & Aggressive NaNs, two novel variants of max pooling and unpooling that identify numerically unstable voxels and replace them with NaNs, allowing subsequent layers to skip computations on irrelevant data. Both methods are implemented within PyTorch and require no architectural changes. We evaluate these approaches on four CNN models spanning neuroimaging and image classification tasks. For inputs containing at least 50% NaNs, we observe consistent runtime improvements; for data with more than two-thirds NaNs )common in several neuroimaging settings) we achieve an average inference speedup of 1.67x. Conservative NaNs reduces convolution operations by an average of 30% across models and datasets, with no measurable performance degradation, and can skip up to 64.64% of convolutions in specific layers. Aggressive NaNs can skip up to 69.30% of convolutions but may occasionally affect performance. Overall, these methods demonstrate that numerical uncertainty can be exploited to reduce redundant computation and improve inference efficiency in CNNs.

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.