Unpaired Translation of Chest X-ray Images for Lung Opacity Diagnosis via Adaptive Activation Masks and Cross-Domain Alignment

TL;DR

This work tackles the challenge of lung opacity in chest X-rays by proposing an unpaired translation framework that removes opacities while preserving anatomical and radiomic features. It introduces adaptive activation masks (AASPM) to guide selective region modification and a cross-domain alignment module (CDAM) that leverages a frozen pre-trained CXR classifier for feature- and label-level consistency. The method achieves superior translation quality (lower FID and KID) and enhances downstream tasks such as lung border segmentation and lesion classification across RSNA, MIMIC-CXR-ARDS, and JSRT datasets; ablation confirms the contribution of each component. The results suggest practical clinical value by improving visualization, segmentation reliability, and opacity detection, with future plans to reduce artifacts and validate in out-of-distribution settings through prospective studies.

Abstract

Chest X-ray radiographs (CXRs) play a pivotal role in diagnosing and monitoring cardiopulmonary diseases. However, lung opacities in CXRs frequently obscure anatomical structures, impeding clear identification of lung borders and complicating the localization of pathology. This challenge significantly hampers segmentation accuracy and precise lesion identification, which are crucial for diagnosis. To tackle these issues, our study proposes an unpaired CXR translation framework that converts CXRs with lung opacities into counterparts without lung opacities while preserving semantic features. Central to our approach is the use of adaptive activation masks to selectively modify opacity regions in lung CXRs. Cross-domain alignment ensures translated CXRs without opacity issues align with feature maps and prediction labels from a pre-trained CXR lesion classifier, facilitating the interpretability of the translation process. We validate our method using RSNA, MIMIC-CXR-JPG and JSRT datasets, demonstrating superior translation quality through lower Frechet Inception Distance (FID) and Kernel Inception Distance (KID) scores compared to existing methods (FID: 67.18 vs. 210.4, KID: 0.01604 vs. 0.225). Evaluation on RSNA opacity, MIMIC acute respiratory distress syndrome (ARDS) patient CXRs and JSRT CXRs show our method enhances segmentation accuracy of lung borders and improves lesion classification, further underscoring its potential in clinical settings (RSNA: mIoU: 76.58% vs. 62.58%, Sensitivity: 85.58% vs. 77.03%; MIMIC ARDS: mIoU: 86.20% vs. 72.07%, Sensitivity: 92.68% vs. 86.85%; JSRT: mIoU: 91.08% vs. 85.6%, Sensitivity: 97.62% vs. 95.04%). Our approach advances CXR imaging analysis, especially in investigating segmentation impacts through image translation techniques.

Figures (6)

• Figure 1: Challenges identified in both lung segmentation with lung opacities and the unpaired image translation approach. 1) Difficulty in correctly segmenting lung regions given CXRs with lung opacities. 2) Interpretability of the translation process when unexpected translation results occur. 3) Introducing unnecessary noise and artifacts. The C1, C2, C3 indicate the numbering from the contributions.
• Figure 2: Top: Model Overview of Proposed Method.To eliminate the opacity parts in the CXRs, both CXRs with and without lung opacities were trained in a cycle using two domains and parameters of the generators that were learned for each direction. During the inference stage, only Generator A is used to generate the output of the CXRs. Bottom Right:Illustration of Cross-Domain Alignment Module (CDAM). Two types of alignment are present in this module: feature alignment and label consistency alignment. Bottom Left:Graphical Illustration of AASP Module. Two types of activation mask penalties and one minimization loss are used to control the generation of the activation masks. Intra-domain translations refer to cases where the input CXR domain matches the generator's output domain (e.g., A $\rightarrow$ A), whereas inter-domain translations occur when the input CXR domain differs from the generator’s output domain (e.g., A $\rightarrow$ B).
• Figure 3: Qualitative Comparison of Translated CXRs Generated by Various Methods. Each row (a) through (c) represents two examples of CXR. Columns from left to right show the original CXR, our method, Munit, Unit, Drit, CycleGAN, and Uvcgan. CXRs without lung opacities generated by our method show better-translated image quality in terms of appropriate pixel intensity, clear lung borders, and the preservation of necessary details compared to the other methods.
• Figure 4: Qualitative Comparison of Segmentation Results on Lung CXRs with Opacities and Translated Lung CXRs without Opacities. Each row (a) through (c) represents three different CXR images. Columns from left to right show: Expert Segmentation Mask, CXR with Opacification, Segmentation of CXR with Opacification, Translated CXR without Opacification, and Segmentation of Translated CXR without Opacification. After the translation, the regions of lung opacities are removed, enabling more accurate lung segmentations given the same lung segmentation model.
• Figure 5: Comparative Analysis of Translation Results and Activation Masks. Rows (a) through (c) present distinct chest X-ray (CXR) examples. Activation masks generated by the model correctly delineate regions of transformation during translation, providing a mechanism to interpret the process of translation during the removal of lung opacities.