IntrinsicX Data Pipeline: reproducible multi-source harmonization of chest X-ray, CT, and brain MRI for medical image synthesis

Reproducible software pipeline that builds the IntrinsicX corpus from five public medical imaging archives: NIH ChestX-ray14, NIH DeepLesion, BraTS-Africa, LIDC-IDRI, and OpenBHB. The pipeline performs source-specific decoding and physically correct intensity restoration, including reversal of the Hounsfield-Unit storage offset and per-image DICOM windowing for computed tomography and per-slice percentile normalization for magnetic resonance imaging; occupancy-based informative axial slice extraction from volumetric sources; resampling to a common 256x256 grid; strictly patient-level 70/15/15 partitioning under a fixed seed; and serialization into a single HDF5 schema. It additionally estimates healthy-tissue reference statistics by unsupervised four-class intensity clustering of confirmed-healthy slices, yielding plausibility-loss targets for brain MRI and chest radiography. Applied to the five source archives, the pipeline produces 841,739 harmonized 256x256 images spanning three modalities and 57 real acquisition sites. This repository contains code only. It regenerates the corpus from the original public archives and does not redistribute imaging data; each source must be obtained under its own access terms. The resulting dataset is deposited separately at doi:10.17632/6m4kb3vffg.2. The IntrinsicX-GAN model code is not part of this release and will accompany a separate modelling publication.