AnyCXR: Human Anatomy Segmentation of Chest X-ray at Any Acquisition Position using Multi-stage Domain Randomized Synthetic Data with Imperfect Annotations and Conditional Joint Annotation Regularization Learning

summary

The gist

Robust anatomical segmentation of chest X-rays (CXRs) remains challenging due to the scarcity of comprehensive annotations and the substantial variability of real-world acquisition conditions.

In short

AnyCXR is a framework that segments organs in chest X-rays across any angle using only synthetic data and imperfect labels. It uses a Multi-stage Domain Randomization engine to create 100,000 diverse training images from CT scans. A learning strategy called Conditional Joint Annotation Regularization enforces anatomical consistency in the model's predictions, allowing it to generalize well to real-world X-rays.

Key concepts

Multi-stage Domain Randomization (MSDR)
This engine generates over 100,000 synthetic chest X-rays from 3D CT volumes. It randomizes image properties like Hounsfield Units and artifacts during pre-projection and then modifies the geometry during in-projection to simulate diverse real-world acquisition conditions, ensuring the model learns robust features.
Conditional Joint Annotation Regularization (CAR)
This is a learning strategy used to handle partial or imperfect labels. It uses a latent space regularization technique called Distribution Loss (LDist) to force the model's predicted organ configurations to match anatomically plausible layouts, ensuring consistency even when training data is incomplete.
Synthetic Supervision
Instead of relying solely on scarce real-world annotations, this method trains the segmentation model entirely on synthetic images generated from 3D CT volumes. This allows the framework to learn generalizable anatomical features that are not tied to specific, limited real-world labeling schemes.
Zero-shot Generalization
This refers to the model's ability to accurately segment structures in novel, unseen real-world datasets without any additional training on those specific datasets. AnyCXR demonstrates strong zero-shot performance across multiple real CXR datasets by leveraging the diverse synthetic training.

Terminology used across episodes

This episode discusses

The paper

AnyCXR: Human Anatomy Segmentation of Chest X-ray at Any Acquisition Position using Multi-stage Domain Randomized Synthetic Data with Imperfect Annotations and Conditional Joint Annotation Regularization Learning · Read on arXiv

Zifei Dong, Wenjie Wu, Jinkui Hao, Tianqi Chen, Ziqiao Wenga, Bo Zhou

Department of Radiology, Northwestern University · Data Science Institute, Vanderbilt University Department of Orthopedics, The Second Hospital of Shanxi Medical University Second Clinical Medical College, Shanxi Medical University School of Computer Science, University of Sydney

Robust anatomical segmentation of chest X-rays (CXRs) remains challenging due to the scarcity of comprehensive annotations and the substantial variability of real-world acquisition conditions. We propose AnyCXR, a unified framework that enables generalizable multi-organ segmentation across arbitrary CXR projection angles using only synthetic supervision. The method combines a Multi-stage Domain Randomization (MSDR) engine, which generates over 100,000 anatomically faithful and highly diverse synthetic radiographs from 3D CT volumes, with a Conditional Joint Annotation Regularization (CAR) learning strategy that leverages partial and imperfect labels by enforcing anatomical consistency in a latent space. Trained entirely on synthetic data, AnyCXR achieves strong zero-shot generalization on multiple real-world datasets, providing accurate delineation of 54 anatomical structures in PA, lateral, and oblique views. The resulting segmentation maps support downstream clinical tasks, including automated cardiothoracic ratio estimation, spine curvature assessment, and disease classification, where the incorporation of anatomical priors improves diagnostic performance. These results demonstrate that AnyCXR establishes a scalable and reliable foundation for anatomy-aware CXR analysis and offers a practical pathway toward reducing annotation burdens while improving robustness across diverse imaging conditions.

Transcript

Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.

Tom: I'm Tom.

Jane: And I'm Jane, here with Lu, Meng and our in-house large language model, Lalam.

Tom: We're looking at the paper: "AnyCXR: Human Anatomy Segmentation of Chest X-ray at Any Acquisition Position using Multi-stage Domain Randomized Synthetic Data with Imperfect Annotations and Conditional Joint Annotation Regularization Learning".

Jane: Here's the gist of the paper. Robust anatomical segmentation of chest X-rays (CXRs) remains challenging due to the scarcity of comprehensive annotations and the substantial variability of real-world acquisition conditions.

Tom: AnyCXR is a unified framework that enables generalizable multi-organ segmentation across arbitrary CXR projection angles using only synthetic supervision.

Jane: The Core Problem Addressed The paper addresses the critical limitation in current CXR segmentation where no generalizable or robust framework exists for delineating comprehensive anatomical structures across diverse acquisition conditions.

Tom: Fundamental obstacles include the lack of large-scale, consistently well-annotated CXR datasets and the inherent ambiguity and incompleteness of 2D manual annotations.

Paper summary: Jane: To follow up on that, how exactly does this AnyCXR framework work? What’s the thesis they are pushing here about how they tackle those challenges?

Tom: The core thesis is generating over one hundred thousand anatomically faithful and highly diverse synthetic radiographs from three dee CT volumes using a Multi-stage Domain Randomization engine >

Lu: That engine does two main things, first, it has this segmentation-driven data curation to ensure anatomical plausibility and label reliability through a two-stage Quality Control procedure >

Meng: Then, they use a multi-stage randomized projection process that modifies Hounsfield Units and image artifacts before simulating the detector response post-projection >

Jane: And then they take this curated dataset and build the Conditional Joint Annotation Regularization framework to enable effective training from incomplete supervision >

Tom: They extend joint regularization learning to multi-anatomical segmentation by transferring features into a latent representation and imposing conditional consistency constraints across overlapping anatomical structures >

Lalam: This allows the model to leverage all available data, meaning both complete and partial annotations, while maintaining inter-organ coherence >

Conclusion: Tom: So we're wrapping up with this AnyCXR paper, which is about human anatomy segmentation for chest X-rays no matter how they were taken >

Jane: Right, so we’ve seen how they used synthetic data and this regularization technique to handle imperfect labels, which really tackles that annotation scarcity problem >

Tom: It seems like the main idea here is making one model that works well across all those different ways the X-rays can be projected using this domain randomization approach >

Lu: I think the real strength is combining that massive multi-stage domain randomization engine with the conditional joint annotation regularization learning strategy to achieve this generality >

Meng: From a deployment standpoint, it’s cool they trained it entirely on synthetic data so they don't need to spend all their time chasing perfect annotations right now >

Lalam: I'm pretty sure this work shows that you can build a robust foundation for anatomy-aware chest X-ray analysis without needing millions of perfectly labeled real images first >

Jane: It really feels like they’ve built a scalable way to reduce the burden of getting those expensive, perfect annotations for every single projection angle >

Tom: The authors are Akhter Singh Vatsa and others, and they’ve shown that this framework can delineate fifty-four anatomical structures in PA, lateral, and oblique views quite accurately >

Lu: And the results on real data show strong zero-shot generalization on multiple different real-world datasets which is a big deal for moving this past just synthetic testing >

Meng: So what does this mean practically for someone trying to set up a screening system? Can they actually rely on this model without retraining it every time they get a slightly different patient setup?

Jane: It suggests that AnyCXR provides those anatomically informed features that help with things like estimating the cardiothoracic ratio or assessing spine curvature >

Tom: And for disease classification, the segmentation-guided model showed an absolute increase of about two point zero four percent over a baseline, which is a solid number to look at > <ref:2512.17263#pg1>

Lalam: I’m confident this approach establishes a reliable foundation for population-level screening workflows by making the analysis more robust across diverse imaging conditions >

Jane: It definitely sets up the next question about how physically realistic those synthetic projections need to be to truly match what happens in the real world.

More episodes

← Home