ART-ASyn: Anatomy-aware Realistic Texture-based Anomaly Synthesis Framework for Chest X-Rays

TL;DR

This work tackles the lack of pixel-level anomaly localization in chest X-rays by proposing ART-ASyn, an anatomy-aware texture-based anomaly synthesis framework. It combines PBTSeg, a progressive, unsupervised lung segmentation method, with a texture-based synthesis pipeline to generate anatomically plausible lung-opacity anomalies and exact pixel-level masks, enabling explicit supervision. The model uses feature-alignment and region-focused losses to minimize reconstruction errors in normal regions while enhancing abnormal localization, and it is evaluated in a zero-shot setting on unseen data, achieving strong image-level and notably better pixel-level segmentation performance than prior methods. The approach demonstrates solid generalization, realistic anomaly synthesis, and interpretable localization, with practical implications for scalable, annotation-efficient chest X-ray anomaly detection.

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 Anatomy-aware Realistic Texture-based Anomaly Synthesis 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 Progressive Binary Thresholding Segmentation 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.

Figures (4)

• 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.
• Figure 2: (a) PBTSeg process illustration of lung mask segmentation using multi-thresholding, where the second row shows candidate regions after rule-based selection for each threshold, and the last row illustrates the iterative process of updating the segmentation regions; (b) Comparison of lung segmentation results using our PBTSeg method against the threshold-based approach adopted in AnatPaste for CXR images.
• Figure 3: (a) Comparison of anomaly synthesis methods, for each method, the first column illustrates the anomaly synthesized and the second column highlights the anomaly area in red; (b) Comparison of image-level reconstruction, for each method, the first column illustrates the reconstructed image and the second column shows the pixel-wise difference $A$ in heatmap format.
• Figure 4: Comparison of anomaly segmentation on the QaTa-ZeroShot dataset.