Going Beyond ImageNet Preprocessing for X-ray Vision Models
Building a CR image preprocessor from the ground up
X-rays are not natural images, so their preprocessing shouldn't be borrowed from ImageNet. We measure what makes CR images different, build a pipeline for them, and ablate each design choice.
Abstract
Most vision models for radiology are fed by a preprocessing pipeline inherited from natural images: normalize the pixels, crop or squeeze the frame into a fixed square, and hand it to the backbone. X-ray images are not natural images, however, and the preprocessing pipeline should reflect this. In this post, we identify and measure a handful of image characteristics specific to CR (computed radiography) and propose our CR image preprocessing pipeline. We run a series of ablation experiments to study the sensitivity of pathology classification models to the image preprocessor's design decisions.
Cite
@misc{karpur2026beyond,
title = {Going Beyond ImageNet Preprocessing for X-ray Vision Models},
author = {Arjun Karpur and Aakash Mishra},
year = {2026},
howpublished = {Epsilon Health Research},
url = {https://www.epsilon.health/research/cr-image-preprocessing}
}