[논문 리뷰] XLSor: A Robust and Accurate Lung Segmentor on Chest X-Rays Using Criss-Cross Attention and Customized Radiorealistic Abnormalities Generation

Youbao Tang, Yuxing Tang외 2명|arXiv (Cornell University)|2019. 04. 19.

COVID-19 diagnosis using AI참고 문헌 31인용 수 86

한 줄 요약

XLSor introduces criss-cross attention for global context in lung segmentation on chest X-rays and uses radiorealistic abnormal CXR data augmentation to improve robustness, achieving superior performance on abnormal and normal CXRs.

ABSTRACT

This paper proposes a novel framework for lung segmentation in chest X-rays. It consists of two key contributions, a criss-cross attention based segmentation network and radiorealistic chest X-ray image synthesis (i.e. a synthesized radiograph that appears anatomically realistic) for data augmentation. The criss-cross attention modules capture rich global contextual information in both horizontal and vertical directions for all the pixels thus facilitating accurate lung segmentation. To reduce the manual annotation burden and to train a robust lung segmentor that can be adapted to pathological lungs with hazy lung boundaries, an image-to-image translation module is employed to synthesize radiorealistic abnormal CXRs from the source of normal ones for data augmentation. The lung masks of synthetic abnormal CXRs are propagated from the segmentation results of their normal counterparts, and then serve as pseudo masks for robust segmentor training. In addition, we annotate 100 CXRs with lung masks on a more challenging NIH Chest X-ray dataset containing both posterioranterior and anteroposterior views for evaluation. Extensive experiments validate the robustness and effectiveness of the proposed framework. The code and data can be found from https://github.com/rsummers11/CADLab/tree/master/Lung_Segmentation_XLSor .

연구 동기 및 목표

• Motivate accurate lung segmentation for chest X-rays to aid automated diagnosis across normal and abnormal cases.
• Develop a segmentation framework that captures global contextual information using criss-cross attention (CCA).
• Reduce annotation burden via radiorealistic abnormal CXR generation and pseudo-masks for robust training.
• Evaluate on multiple datasets including NI H Chest X-rays with PA and AP views to test generalization.

제안 방법

• Use criss-cross attention modules within a CNN backbone to capture horizontal and vertical global context for pixel-wise segmentation.
• Replace last down-sampling layers with dilated convolutions to achieve a stride of 8 for high-resolution context.
• Adopt a two-CCA recurrent arrangement to densely collect contextual information.
• Generate radiorealistic abnormal CXRs from normal ones via MUNIT and propagate normal masks as pseudo masks for training augmentation.
• Train XLSor with real data and augmented data; evaluate against U-Net baselines.
• Annotate 100 abnormal NIH CXRs to test generalization.

실험 결과

연구 질문

• RQ1 Can criss-cross attention improve global contextual reasoning for accurate lung segmentation in CXRs, especially under abnormal/pathological conditions?
• RQ2 Does radiorealistic augmentation with pseudo masks boost robustness and generalization to unseen abnormal CXRs across PA and AP views?
• RQ3 How does XLSor compare to a standard U-Net baseline in normal and abnormal CXR segmentation settings?

주요 결과

CXR 마스크	U-Net R	U-Net A^4	U-Net R+A^4	XLSor R	XLSor A^4	XLSor R+A^4
REC	0.976 ± 0.02	0.963 ± 0.03	0.972 ± 0.02	0.973 ± 0.02	0.967 ± 0.02	0.974 ± 0.02
PRE	0.968 ± 0.03	0.979 ± 0.02	0.970 ± 0.03	0.979 ± 0.02	0.983 ± 0.01	0.976 ± 0.01
DICE	0.972 ± 0.02	0.971 ± 0.02	0.975 ± 0.01	0.976 ± 0.01	0.973 ± 0.01	0.975 ± 0.01
AVD	0.198 ± 0.56	0.162 ± 0.36	0.131 ± 0.34	0.149 ± 0.51	0.098 ± 0.07	0.078 ± 0.06
VS	0.988 ± 0.02	0.989 ± 0.01	0.990 ± 0.01	0.992 ± 0.01	0.991 ± 0.01	0.993 ± 0.01

• XLSor outperforms U-Net on both public testing and NIH datasets, with notable gains on unseen abnormal CXRs (e.g., Dice score improvements).
• Adding radiorealistic augmented samples (A^i) consistently improves performance over using real data alone.
• Using criss-cross attention with CCA modules enhances global context learning and yields better segmentation, especially in challenging cases.
• Models trained with augmentation (R+A^4) achieve robust performance on NIH data, indicating good generalization to more complex abnormal CXRs.
• Even when trained only on augmented data (A^4), XLSor achieves competitive results on public and NIH datasets, demonstrating effective pseudo-mask supervision.