Lazy Resampling: Fast and information preserving preprocessing for deep learning.

Area of Science:

• Deep learning
• Computer vision
• Medical imaging

Background:

• Data preprocessing is crucial for deep learning workflows, impacting network stability and generalization.
• Traditional preprocessing pipelines involve multiple resampling stages, leading to increased execution time, image quality degradation, and bias.
• Complex pipelines, especially in medical imaging, can introduce artifacts and hinder effective data manipulation.

Purpose of the Study:

• To introduce Lazy Resampling, a novel software that reframes spatial preprocessing operations.
• To reduce the computational cost and signal degradation associated with multi-stage resampling pipelines.
• To simplify pipeline design and enable non-destructive operations like cropping in medical imaging.

Main Methods:

• Rephrasing spatial preprocessing as a graphics pipeline where transforms generate descriptions.
• Compositing transform descriptions into a single resample operation to minimize individual data modifications.
• Implementing Lazy Resampling to provide benefits without requiring users to alter their pipeline construction methods.

Main Results:

• Demonstrated lower information loss in lazy resampling pipelines compared to traditional ones.
• Showcased the ability of Lazy Resampling to prevent catastrophic loss of semantic segmentation accuracy during label inversion.
• Achieved statistically significant improvements in training UNets for semantic segmentation tasks.

Conclusions:

• Lazy Resampling effectively minimizes information loss in multi-resampling pipelines.
• Enables simpler, non-destructive preprocessing pipelines, improving ease of use for researchers.
• Facilitates accurate label inversion and shows promise for integration into major deep learning libraries like MONAI.