Multimorbidity-Aware Proxy Latent Inference for Report-Conditioned 3D CT Generation

Text-conditioned generation of volumetric computed tomography (CT) from radiology reports is limited by the absence of explicit spatial information. Retrieval-augmented methods provide anatomical guidance, but existing approaches rely on a single nearest-neighbour case, which may not represent multiple co-occurring findings. We propose a multimorbidity-aware proxy latent inference framework centred on spatial confidence-weighted fusion of multiple retrieved proxies. Reports are decomposed into global and finding-level queries to retrieve a diverse candidate set. Semantic relevance and inter-proxy agreement are then combined to infer a spatial anatomical prior, while confidence-coupled gating and anatomy regularisation support its integration into a latent diffusion generator. On CT-RATE, the proposed method improves image fidelity over reproduced RAG-Nearest, reducing average 2.5D FID from 0.306 to 0.271, and improves clinical consistency, increasing macro AUC from 0.792 to 0.812. Gains are largest for reports containing multiple findings. Against a common, method-independent anatomical reference, Dice increases from 0.704 to 0.726 and HD95 decreases from 3.92 to 3.61 mm. Random-pairing and pipeline-ceiling calibration show that these improvements are statistically supported but modest relative to the range attributable to conserved gross anatomy and pipeline reproducibility. The spatial mixture weights also provide regional attribution of the retrieved proxies contributing to each generated volume.