Generative adversarial network–based image translation (GAN-IT) has emerged as a promising approach for chest X-ray (CXR) anomaly localization (AL-CXR) without pixel-level annotations, especially when training relies on unpaired data. However, the nonstandardized, unpaired nature of available datasets impedes learning of discriminative features that separate normal from abnormal cases, yielding unstable and sometimes inaccurate performance. To overcome this limitation, we propose a GAN-IT pipeline that includes an image-registration step to produce pseudo-pairs. Specifically, we introduce a deep learning–based registration module that converts unpaired images into pseudo-paired samples via a two-stage procedure: first, global linear alignment for coarse thoracic normalization; second, an AI-based nonlinear refinement that captures fine, spatially varying deformations. Composing these transforms reduces artifacts inherent to unpaired training and stabilizes translation. We evaluate our method on two public CXR benchmarks (tuberculosis and consolidation) and integrate it into two representative unpaired translation models (CycleGAN and CUT). Using patient-wise anomaly-score AUC, our IT-DPR improves tuberculosis performance from 0.755 to 0.928 and consolidation performance from 0.964 to 0.991, while showing reduced sensitivity to threshold choices and improved stability across operating points. Under identical settings, it also consistently outperforms the registration-integrated baseline Reg-GAN (0.802 and 0.807 on tuberculosis and consolidation, respectively). To clarify clinical relevance in the absence of pixel-level ground truth, we further assess localization using box-level annotations where available. Overall, the proposed registration module is readily compatible with existing GAN-IT architectures and provides consistent gains in both accuracy and stability for AL-CXR in low-annotation settings.
Oh et al. (Thu,) studied this question.