AURAD: Controllable Chest X-ray Synthesis

• AURAD is a controllable chest radiograph synthesis framework that employs a two-stage pipeline to generate images and pseudo semantic masks for precise pathology localization.
• It integrates structured clinical prompts with anatomical masks using InstructPix2Pix and ControlNet modifications, achieving superior results (FID 32.83, Path AUC 59.31).
• The framework enhances downstream detection, segmentation, and classification tasks by augmenting scarce data, even boosting performance under extreme low-resource conditions.

AURAD is a controllable chest X-ray synthesis framework that jointly generates high-fidelity radiographs and pseudo semantic masks through a progressive pipeline in which pseudo masks are first generated from clinical prompts conditioned on anatomical structures, and then used to guide image synthesis. It is explicitly introduced as “Anatomy–Pathology Unified Radiology Synthesis with Progressive Representations” and is designed for chest radiographs, where disease patterns are morphologically diverse and tightly intertwined with anatomical structures (Ding et al., 5 Sep 2025). In the cited arXiv corpus, this is the only resource formally named AURAD; nearby acronyms denote unrelated systems, including AURSAD for robotic screwdriving anomaly detection (Leporowski et al., 2021) and several distinct AURA methods in speech agents, kinodynamic replanning, LLM alignment, LLM-judge auditing, industrial smoke detection, and audio-to-LLM adaptation (Maben et al., 29 Jun 2025, Golestaneh et al., 26 May 2026, Adak et al., 8 Aug 2025, Zhang et al., 18 Jun 2026, Bychkov et al., 1 Aug 2025, Cheng et al., 9 Jun 2026).

1. Problem domain and motivation

AURAD addresses controllable chest radiograph synthesis in a regime where data scarcity, annotation cost, and domain specificity are especially restrictive. The motivating use case is augmentation for downstream medical vision under limited supervision, particularly when dense lesion masks are much rarer than image-level labels. The paper argues that chest X-rays are a difficult synthesis target because pathology boundaries are often ambiguous, multiple diseases can coexist in the same image, disease appearance is tightly constrained by anatomy, and radiographic grayscale texture and projection effects limit transfer from natural-image generation methods (Ding et al., 5 Sep 2025).

The method is positioned against several limitations in prior work. Image-only GAN and diffusion systems can produce visually plausible chest radiographs but offer limited semantic and spatial control. General-purpose text-to-image models are said to suffer from domain shift on grayscale medical imagery and modality-specific terminology. Domain-adapted report-to-CXR models such as RoentGen and Cheff do not explicitly encode anatomy–pathology spatial relations, while bounding-box-guided approaches such as XReal remain coarse for many radiographic findings and usually require manual input. A central target of AURAD is therefore not generic realism alone, but synthesis that is simultaneously realistic, diagnostically meaningful, spatially controllable, and usable for downstream detection and segmentation.

The paper’s specific critique of earlier controllable synthesis is the use of randomly sampled masks or weak disease priors. Those masks are described as limiting pathology diversity, failing to model multi-disease coexistence, and not enforcing the fact that disease patterns should align with organ structures. AURAD’s response is to learn pathology masks conditioned on anatomy and structured prompts rather than sampling pathology occupancy heuristically.

2. Progressive architecture

The defining design choice in AURAD is a two-stage progressive pipeline: text-to-mask generation followed by mask-to-image generation. The first stage converts a structured clinical prompt plus anatomical context into a dense pseudo pathology mask. The second stage uses the generated mask as a structural prior to synthesize the final chest X-ray. The paper describes this as a progression from semantic specification to dense spatial structure and finally to image appearance, with the intermediate mask encoding pathology type, location, extent, and coexistence (Ding et al., 5 Sep 2025).

Stage 1 adapts InstructPix2Pix to generate pathology-augmented segmentation maps from organ-only anatomical masks and disease-description prompts. Because the model receives anatomy as input, pathology placement is constrained by organ geometry rather than free-form image coordinates. Stage 2 extends RoentGen with ControlNet so that the generated pathology-plus-organ mask is encoded through a convolutional branch and injected into RoentGen’s frozen U-Net via zero-convolution layers. The output is a synthetic chest X-ray conditioned both on the prompt and on the spatial structure already established by the mask.

AURAD inserts two explicit quality-control stages between and after generation. First, generated masks undergo Prompt-Match Validation: a custom prompt tool re-captions the mask by analyzing overlap between pathology and organ regions, and masks whose recovered prompt does not match the input prompt are rejected and regenerated. Second, synthesized images are filtered by pretrained expert models, including a lightweight binary discriminator for visual realism and a pretrained multi-label CXR classifier for pathology plausibility and anatomical consistency.

The paper does not explicitly write the diffusion loss for either stage, nor a classifier-free guidance formula, GAN objective, or reconstruction loss. A plausible formalization of the progressive design is

$$p(x, m \mid t, a) = p(m \mid t, a)\, p(x \mid m, t),$$

with $t$ the clinical prompt, $a$ the anatomical mask, $m$ the generated pathology-augmented semantic mask, and $x$ the synthesized chest X-ray. This factorization is consistent with the pipeline, but it is not typeset in the manuscript.

3. Anatomy-aware representation and prompt construction

AURAD’s intermediate representation is built from organ masks produced by a pretrained anatomy segmenter from TorchXRayVision, specifically a PSPNet-based model. The appendix states that the lungs are segmented into left and right lungs, each further divided into upper, middle, and lower regions, yielding six subregions and five major regions: left lung, right lung, bilateral lung, mediastinum, and heart (Ding et al., 5 Sep 2025). This provides clinically meaningful spatial anchors for prompt-conditioned pathology synthesis.

Prompts are structured rather than free-form. The format is:

“A photo of a Chest X-ray with [SEVERITY1] [CLASS1] on [LOC1], [SEVERITY2] [CLASS2] on [LOC2], ...”

The prompt vocabulary includes three severity levels—mild, moderate, severe—and eleven locations spanning left/right upper, middle, and lower lung, mediastinum, heart, and left/right/bilateral lung. The paper’s formulation/table uses fourteen disease classes, while the dataset tables cover thirteen abnormal classes plus normal. This structured prompting is presented as easier to learn under limited data than unconstrained text, and as better aligned with the paper’s anatomy-conditioned generation logic.

The prompt tool derives text from disease masks and organ masks by computing pathology–organ overlap. Given disease class $C$, organ mask $O$, and pathology mask $P$, the algorithm partitions $O$ into anatomical parts $O_i$, computes

$t$0

then overlap area

$t$1

selects the dominant overlapping region, derives a location label, and assigns severity from overlap magnitude. Cardiomegaly is handled separately through cardiothoracic ratio thresholds:

$t$2

$t$3

$t$4

The resulting masks are intended to encode multi-pathology coexistence and anatomy–pathology consistency rather than arbitrary occupancy. The paper notes, however, that the prompt tool has limitations: it relies on overlap and does not encode relative ordering among multiple diseases.

4. Data, supervision, and implementation

The training set consists of 3001 images from the ChestX-Det training split plus 1000 randomly selected abnormal pairs from CANDID-PTX, for a total of 4001 images. Evaluation uses the ChestX-Det test set (542 images), Chest X-ray14 test set (741 images), VinDr-CXR test set (2938 images), and MIMIC-CXR test set (2675 images). Supplementary anatomical masks from MIMIC-CXR are also sampled during inference to increase anatomical diversity (Ding et al., 5 Sep 2025).

AURAD does not require paired report–mask–image triples. Pathology masks come from ChestX-Det and CANDID-PTX; anatomy masks are extracted automatically using TorchXRayVision; prompts are constructed rule-based from pathology–organ overlap and severity heuristics; and a pretrained TorchXRayVision classifier is used for verification. This reduces dependence on manually authored multimodal annotations while preserving structured controllability.

Preprocessing is fixed to frontal PA/AP views, with the shortest edge resized to 512 and center-cropped to 512 × 512. The implementation uses 4 × NVIDIA A100 80GB, PyTorch, TorchXRayVision, and Diffusers. Both generation stages are trained for 6000 steps with learning rate $t$5 and batch size 128 total / 32 per GPU; wall-clock time is about 7 h per stage. At inference, Stage 1 uses UniPCMultistep with 20 steps and takes about 5 s, while Stage 2 uses UniPCMultistep with 70 steps and takes about 11 s. With self-assessment and retries, inference averages about 16 s per attempt and about 60 s per sample, with about 3–4 attempts on average and worst-case latency scaling linearly with max_attempts.

The paper states that code, pretrained models, and the synthesized dataset will be released upon publication, but several details remain unspecified, including exact ControlNet layer configuration, exact mask-channel encoding, classifier/discriminator thresholds, and the precise synthetic-to-real sampling curriculum beyond the reported ratios.

5. Empirical performance and downstream utility

AURAD is evaluated both as an image-synthesis system and as a generator of useful supervision for downstream models. In the main image-generation comparison against Cheff and RoentGen, the reported results are:

Method	FID	MS-SSIM	Path AUC
Cheff	59.37	0.4290	52.34
RoentGen	56.39	0.4031	56.51
AURAD	32.83	0.5968	59.31

The corresponding CLIP scores are 30.64 for Cheff, 29.65 for RoentGen, and 30.14 for AURAD; a real-data pathology AUC baseline of 57.64 is also reported (Ding et al., 5 Sep 2025). Within the paper’s evaluation protocol, this means AURAD achieves the best FID, the best structural consistency, and the highest pathology recognizability, while remaining comparable in text–image semantic alignment.

The radiologist study uses 3 board-certified radiologists with 5–15 years of thoracic imaging experience. In a blinded realism study on 20 patients, 78% of AURAD images were judged realistic or authentic after thresholding scores $t$6 on a 1–5 Likert scale. In a separate study on 50 patients, 41% of segmentation overlays predicted by a downstream Mask R-CNN trained with synthetic augmentation were judged clinically useful. Reported inter-rater agreement is moderate, with realism ICC = 0.40 and usefulness Fleiss’ $t$7 to 0.53 depending on section.

AURAD’s central applied claim is that synthesized image–mask pairs improve downstream performance. Using Mask R-CNN with FPN trained from scratch, augmentation with AURAD improves detection mAP@50 on ChestX-Det from 43.65 to 46.94, on VinDr-CXR from 20.39 to 21.79, and on ChestX-ray14 from 26.93 to 31.96. On ChestX-Det segmentation, Dice improves from 37.83 to 39.16. On MIMIC-CXR classification, AUC improves from 65.99 to 69.77. Under the most extreme low-resource setting, using only 1% of real data, adding 5x synthetic data increases detection mAP@50 from 7.61% to 25.21%. For targeted long-tail augmentation, atelectasis mAP@50 improves from 43.60 to 48.08, while overall detection increases from 43.65 to 45.28.

These results support the paper’s broader claim that AURAD is not only a realism-oriented generator but also a practical mechanism for producing paired synthetic supervision for detection, segmentation, and classification under scarcity, imbalance, and cross-dataset shift.

6. Ablations and analytical findings

The ablation studies isolate three aspects of the framework: mask quality, anatomical conditioning, and generation backbone choice. In the mask-quality study, real mask, no expert filter yields FID 35.78, MS-SSIM 0.56, Path AUC 56.72, and mAP@50 42.54; synthetic mask, no expert filter improves these to 34.00, 0.57, 57.37, and 44.60; and synthetic mask + expert filter reaches 32.83, 0.60, 59.31, and 46.94 (Ding et al., 5 Sep 2025). The paper interprets this as evidence that learned synthetic masks generalize better than the available radiologist annotations in this setup, and that expert filtering further improves both image quality and downstream utility.

The mask-conditioning ablation is equally central. Removing organ masks reduces performance to FID 56.39, CLIP 29.65, MS-SSIM 0.4031, and Path AUC 56.51. Using only organ masks gives 33.45, 28.93, 0.5709, and 55.96. Full disease-plus-organ conditioning gives 32.83, 30.14, 0.5968, and 59.31. This shows that anatomical conditioning contributes most of the structural gain, while adding disease masks restores and improves pathology alignment.

The paper also studies prompt form and backbone selection. A free-text prompt analysis reports that structured prompts remain best: the structured baseline gives FID 32.83 and Path AUC 59.31, while end-to-end free-form prompting degrades to FID 37.40 and Path AUC 57.64. For Stage 1, InstructPix2Pix is compared with DiffEdit and Add-It for CXR mask generation. InstructPix2Pix performs best with IoU 0.6897, Dice 0.7772, FID 15.19, MS-SSIM 0.9286, and runtime about 5 s; DiffEdit is reported as much worse with spurious edits, and Add-It as much worse and very slow at about 3 min. A t-SNE analysis on ResNet-18 features further shows that synthetic masks cluster by pathology in a manner comparable to real masks, sometimes more distinctly.

7. Limitations, scope, and nomenclature

The authors state three principal limitations: the method assumes standardized labels across datasets, it struggles with subtle or fine-grained findings such as nodules and fractures, and it does not yet model complex disease interactions (Ding et al., 5 Sep 2025). The appendix and discussion imply additional constraints: performance depends on the quality of pretrained organ segmentation, inaccurate anatomy masks can propagate spatial errors, synthetic masks may be clinically plausible yet disagree with benchmark labels, and text-to-mask generation is more sensitive to prompt formulation than the mask-to-image stage. Free-form prompts are supported, but structured prompts remain superior.

These limitations are directly tied to the paper’s scope. AURAD is strongest when the goal is anatomy-aware, pathology-aware synthesis that produces paired images and masks for downstream tasks. It is less a general-purpose radiology generator than a domain-specific framework for chest X-ray augmentation under scarce dense supervision. This suggests its most natural applications are low-resource detection and segmentation, rare-class balancing, and controllable multi-pathology synthesis.

Within the broader arXiv literature, the term AURAD should be distinguished sharply from similarly named resources. AURSAD denotes the “Universal Robot Screwdriving Anomaly Detection Dataset,” a multivariate time-series benchmark for robotic assembly anomaly detection rather than medical synthesis (Leporowski et al., 2021). Multiple unrelated papers define AURA in speech-driven tool use, kinodynamic replanning, affordance-aware LLM alignment, LLM-judge auditing, industrial smoke detection, and audio-internalized LoRA adaptation (Maben et al., 29 Jun 2025, Golestaneh et al., 26 May 2026, Adak et al., 8 Aug 2025, Zhang et al., 18 Jun 2026, Bychkov et al., 1 Aug 2025, Cheng et al., 9 Jun 2026). In strict nomenclature, AURAD refers to the chest-radiograph synthesis framework alone.