Color variation across histopathology slides remains one of the most stubborn sources of noise in computational pathology. Different scanners, staining protocols, and laboratories produce images that look systematically different even when the underlying tissue is similar. This batch effect degrades the performance of AI models, especially foundation models trained across multiple institutions.

Classical stain normalization methods (Histogram Matching, Macenko, Reinhard) are training-free, interpretable, and effective. The problem is speed. Existing implementations process images sequentially or fail to fully exploit modern GPUs, making them impractical for the tens of thousands of tiles that make up large whole-slide image cohorts.

Training models on radiological images is still slowed by a basic infrastructure problem. Public datasets differ in file formats, directory structures, label schemas, and annotation types. Integrating them into a single training pipeline usually requires custom, one-off engineering for each source.

RadHarmony is an open-source library designed to remove that friction. It standardizes metadata from 24 public datasets into a unified format and delivers samples through a single interface that supports classification labels, segmentation masks, bounding boxes, and radiology reports. The system is built to work with chest radiographs first, with early support for CT and MRI.

Medical AI models can leak sensitive patient information through privacy attacks, even when average risk metrics look acceptable. This paper shows that while aggregate membership inference attack success often appears close to random guessing, individual patient-level risks can be extremely high— with some patients facing near-perfect identifiability.

The risks are also disparate: they disproportionately affect underrepresented groups (by race, disease prevalence, or insurance status). Larger models tend to increase these individual and group-level privacy vulnerabilities.