ART-ASyn: Anatomy-aware Realistic Texture-based Anomaly Synthesis Framework for Chest X-Rays
Qinyi Cao, Jianan Fan, Weidong Cai · 2025
Read ART-ASyn: Anatomy-aware Realistic Texture-based Anomaly Synthesis Framework for Chest X-Rays by Qinyi Cao, Jianan Fan, Weidong Cai (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2512-00310.txt.
Introduction
Finding disease in chest X-rays without labeled examples is the goal of unsupervised anomaly detection. Qinyi Cao, Jianan Fan, and Weidong Cai present ART-ASyn in 2025, a framework that synthesizes realistic anomalies for training. Synthetic anomalies come with known masks, which enables explicit supervision.
Older synthesis methods often produce anomalies that look unlike real pathology and ignore anatomical structure. ART-ASyn is anatomy-aware and texture-based, aiming for irregularities that blend with real tissue patterns. More realistic fake lesions could train detectors that miss fewer real ones, a direct concern for hospital screening programs. Chest X-rays are among the most common medical images, so gains here scale widely. Anatomy-aware synthesis respects ribs and lung boundaries that naive fakes ignore.
Catalog record
| CREATOR | Qinyi Cao, Jianan Fan, Weidong Cai |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 34 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2512-00310
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
ART-ASyn: Anatomy-aware Realistic Texture-based Anomaly Synthesis Framework for Chest X-Rays Qinyi Cao, Jianan Fan, Weidong Cai Abstract Unsupervised anomaly detection aims to identify anomalies without pixel-level annotations. Synthetic anomaly-based methods exhibit a unique capacity to introduce controllable irregularities with known masks, enabling explicit supervision during training. However, existing methods often produce synthetic anomalies that are visually distinct from real pathological patterns and ignore anatomical structure. This paper presents a novel A natomy-aware R ealistic T exture-based A nomaly Syn thesis framework ( ART-ASyn ) for chest X-rays that generates realistic and anatomically consistent lung opacity related anomalies using texture-based augmentation guided by our proposed P rogressive B inary T hresholding Seg mentation method ( PBTSeg ) for lung segmentation. The generated paired samples of synthetic anomalies and their corresponding precise pixel-level anomaly mask for each normal sample enable explicit segmentation supervision. In contrast to prior work limited to one-class classification, ART-ASyn is further evaluated for zero-shot anomaly segmentation, demonstrating generalizability on an unseen dataset without target-domain annotations. Code availability is available at https://github.com/angelacao-hub/ART-ASyn . 1 Introduction Detecting anomalous regions in chest X-ray (CXR) images is critical for clinical decision-making, as these regions often correspond to pathologies requiring timely intervention [ 14 , 22 ] . Beyond image-level classification, precise localization enhances diagnostic interpretability, supports downstream tasks such as treatment planning, and reduces the risk of oversight [ 21 , 7 , 36 , 37 ] . However, in CXR images, precise annotation can be expensive and time-consuming to obtain [ 10 , 19 ] . Hence, detecting anomalous regions without requiring pixel-level ground-truth annotation is highly desirable, enabling scalable and cost-effective deployment of anomaly detection models in clinical settings [ 9 ] . Unsupervised anomaly detection (UAD) has emerged as a promising solution, aiming to learn the distribution of normal data and identify anomalies as deviations from this learned distribution without requiring annotated anomalies or pixel-level supervision [ 15 , 33 , 34 ] . Reconstruction-based methods include both image and feature reconstruction are widely used for anomaly detection, typically rely on generative models such as autoencoders (AEs) [ 1 , 41 , 42 , 2 , 8 , 40 ] , generative adversarial networks (GANs) [ 26 , 27 ] , or diffusers [ 31 , 32 , 3 ] . These models are trained exclusively on normal images with the objective of reconstructing them accurately, under the assumption that unseen abnormal regions will not be faithfully reconstructed and thus exhibit high reconstruction errors during inference. However, these methods suffer from two notable limitations. First, they are prone to the “identical shortcut” problem [ 38 , 39 ] , where the models simply learn to denoise or replicate the input image regardless of whether the content is normal or abnormal. Second, they face a trade-off between sensitivity and precision [ 5 ] , increasing the level of distortion can amplify the reconstruction error in normal regions, while insufficient distortion can lead to poor modeling of normal appearance, thereby reducing the reliability of the anomaly score. In contrast to reconstruction-based approaches, synthesizing-based methods generate artificial anomalies by manipulating normal images, offering the advantage of known spatial locations for explicit supervision. For instance, CutPaste [ 16 ] relocates patches within the same image to introduce anomalies, while NSA [ 28 ] employs Poisson image editing and parameter tuning to create more natural-looking synthetic anomalies. Despite these advances, such methods often produce visually distinct but anatomically implausible anomalies, neglecting domain-specific structural regularities. Building on this idea, AnatPaste [ 25 ] incorporates threshold-based lung segmentation to constrain anomaly synthesis within lung regions. However, its reliance on a single threshold and a simplistic cut-and-paste strategy fails to account for the appearance of more delicate body structures. Consequently, models trained with such data may overfit to superficial cues and fail to generalize to real-world anomalies [ 17 ] . To overcome these limitations, we propose an A natomy-aware R ealistic T exture-based A nomaly Syn thesis method ( ART-ASyn ) for CXRs which synthesizes realistic lung opacity related anomalies using texture brushes that mimic the appearance of chest vessels, airways, and body tissues under CXRs, producing pathologies that resemble real cases. To ensure anatomical consistency, we introduce a P rogressive B inary T hresholding Seg mentation method ( PBTSeg ) that progressively refines thresholded masks to accurately delineate lung regions and constrain anomaly placement within anatomically plausible areas. To reduce reconstruction errors in normal regions and guide the model to focus on abnormal areas, ART-ASyn leverages pixel-level anomaly masks derived by binarizing the synthesized anomalies, enabling explicit supervision for anomaly segmentation and localized loss functions that operate only within abnormal regions. Building on prior studies on CXRs that focus solely on one-class classification, we extend evaluation to anomaly segmentation by testing our model in a zero-shot setting, trained on CheXpert [ 11 ] and ZhangLab [ 12 ] , and evaluated on the unseen QaTa-COV19-v2 [ 35 ] dataset without any pixel-level annotations from the target domain. Our key contributions are as follows: • We propose ART-ASyn , a novel anomaly synthesis method that generates realistic CXR lung opacity related anomalies guided by PBTSeg , a progressive binary thresholding lung segmentation method to ensure anatomical consistency. • The ART-ASyn framework reduces reconstruction errors in normal regions and guides the model to focus on abnormal areas by utilizing pixel-level anomaly masks for explicit anomaly segmentation supervision and localized reconstruction loss. • We demonstrate the model’s generalizability through zero-shot anomaly segmentation on unseen dataset, without using target-domain annotations. Figure 1 : Overview of the ART-ASyn synthesis pipeline for generating realistic synthetic lung opacity related anomalies in CXR images. Random points within the lung region segmented by PBTSeg are painted using a dynamic brush to form an initial anomaly. A series of transformations are then combined and applied to this anomaly, which is overlaid onto the original CXR image to create a synthetic abnormal image. 2 Related Work For unsupervised anomaly detection (UAD), the most closely related approaches fall into two main categories: reconstruction-based and synthesis-based methods. In addition, there exist embedding-based methods [ 13 , 24 , 23 ] that frame anomaly detection as an out-of-distribution (OOD) problem. However, these methods are primarily designed for image-level classification and lack the ability to provide fine-grained localization of anomalies, so we do not discuss them in detail here. 2.1 Reconstruction-based Methods Reconstruction-based methods assume that models trained exclusively on normal data will produce higher reconstruction errors when presented with abnormal regions. These methods typically follow a corrupt-and-reconstruct paradigm, where the input is deliberately corrupted and then reconstructed to align with the learned distribution of normal data. While reconstruction-based methods have good generalizability to adapt across different imaging modalities, they struggle to distinguish between fine-grained anatomical structures and pathological variations, leading to poor localization of subtle anomalies. Traditional autoencoder-based models [ 1 , 41 ] and GAN-based methods [ 26 , 27 ] laid the foundation for this paradigm but tend to overgeneralize, suffering from the “identical shortcut” issue where abnormal regions remain unreconstructed and appear abnormal, leading to false negatives. With the rise of diffusion models in image synthesis [ 6 ] , they were adapted for anomaly detection