Latent-aligned scalable interpolant transformer for controllable chest X-ray synthesis

Abstract Controllable chest X-ray (CXR) synthesis requires control over both semantic findings and their spatial relationship to thoracic anatomy. Text-conditioned generative models can specify disease labels or report-level descriptions but often provide limited control over pathology location and extent. We introduce LatAlignCXR, which represents anatomical layout, pathology location, and residual appearance within a structured latent space. Text prompts, reference radiographs, and user-defined coordinates are mapped to this representation, from which anatomical, pathology, and thoracic structural priors are decoded to condition a scalable interpolant transformer (SiT). The same representation supports text-, image-, and coordinate-conditioned synthesis. On MIMIC-CXR, LatAlignCXR achieved the highest anatomical Dice (0.6564), the lowest Fréchet distance in three of four feature spaces, and the highest pneumonia-localization consistency (Dice 0.1385, IoU 0.0741). In a blinded radiologist evaluation, LatAlignCXR received mean ratings of 4.63 for pathology consistency and 4.50 for spatial consistency, compared with 3.63 and 3.37 for Cheff, respectively. In downstream classification, synthetic augmentation increased accuracy from 69.0% to 76.0% with 1000 real images, with smaller and non-monotonic effects as real-data availability increased. These findings support structured latent conditioning for spatially controllable CXR synthesis, particularly for pathology representation and spatial consistency.