Organ-Conditioned Adversarial Synthesis of 3D CT Volumes from Single-View Frontal Chest X-rays

Medical imaging today faces a structural asymmetry: the frontal chest X-ray (CXR) is the most widely performed radiological examination globally — inexpensive, fast, and low in radiation dose — yet it collapses a volumetric three-dimensional anatomy into a single two-dimensional projection, permanently discarding depth information critical for diagnosis. Computed tomography (CT), by contrast, provides the gold standard three-dimensional anatomical view but at 20–200× higher radiation exposure, significantly higher cost, and limited availability in low-resource clinical environments. This project proposes and trains haloCT-GAN, a novel generative adversarial network architecture that learns to synthesise anatomically coherent 128³-voxel CT volumes from a single frontal chest X-ray, bridging this fundamental imaging gap. 

The core technical challenge is not merely upsampling or hallucinating missing information, but doing so in a way that is anatomically consistent: the predicted CT must reproduce organ shapes, bone structures, vascular pathways, and tissue-density distributions plausible for a real patient — not an average phantom. To achieve this, haloCT-GAN introduces several architectural innovations. The generator employs a dual-branch 2D encoder (DuoLift) to extract rich multi-scale features from the input X-ray, followed by a fully-connected-layer-free bottleneck that reshapes spatial feature maps into volumetric tensors without discarding spatial structure. A 3D U-Net decoder with three Attention Gate skip connections then progressively reconstructs the 3D volume at resolutions of 32³, 64³, and 128³, with the attention mechanism learning to suppress irrelevant anatomical regions and focus on organ boundaries and structural edges. The total generator parameter count is 25.8M — compact relative to the task complexity. 

The Phase 3 discriminator introduces an organ-conditioned adversarial training paradigm: rather than asking the generic question "is this volume real or fake?", the discriminator receives the CT volume concatenated with a 15-class one-hot organ segmentation mask (produced by TotalSegmentator, Apache 2.0), effectively asking per-organ questions — "does this cardiac region look like a real heart?", "do these ribs exhibit realistic cortical bone intensity?". This shifts the adversarial game from global realism to organ-specific anatomical fidelity, a key innovation for clinically meaningful synthesis. 

The loss function architecture is equally multi-scale: a 3D Sobel edge loss (λ=15.0) enforces sharpness at organ boundaries and bone interfaces; a 15-class multi-organ Dice loss (λ=3.0) optimizes volumetric overlap of each anatomical structure; per-organ discriminator feature matching (λ=1.0) provides perceptual-level supervision per tissue type; and an organ-aware texture loss (λ=2.0) constrains tissue-density statistics within each segmented region. These losses are combined with reconstruction (L1, 3D gradient difference), frequency-domain (Log Focal Frequency Loss), and structural (Hessian matching, Hessian eigenvalue) objectives. 

The training dataset is the National Lung Screening Trial (NLST), a high-quality publicly available clinical dataset containing 3,419 paired CT volumes and digitally reconstructed radiographs (DRRs used as X-ray proxies). Phase 3 preprocessing involves running TotalSegmentator inference on all 3,419 volumes to generate the per-volume 15-class organ mask needed for organ-conditioned training. On the current hardware (single NVIDIA GB10), this preprocessing alone requires approximately 57 hours, and training 100 epochs requires an additional ~25 hours — a total of roughly 3.5 days on a single GPU with no margin for ablation studies or dataset expansion. Expanding the dataset to the full NLST cohort (~28,000 volumes) — a critical step for generalisation — would require approximately 20 days of continuous preprocessing on the current hardware, making iterative research cycles practically infeasible. 

This project requests supercomputer access to overcome the compute bottleneck, enable multi-GPU distributed training at Phase 3 scale, and support the ablation studies and hyperparameter searches required to rigorously validate organ-conditioned CT synthesis. Preliminary results from Phase 2 (before organ conditioning) already demonstrate MAE=0.0926 and SSIM=0.8472 on the validation set — confirming the architecture's viability — with medical expert evaluation confirming correct general anatomical structure and visible nodule localisation. Phase 3 is designed to close the remaining quality gaps: sharpened bone detail, defined organ boundaries, and visible vascular structures.