Quality-Gated CycleGAN with Frequency-Split Composition for Structure-Constrained Contrast-Like Enhancement of Non-DICOM Chest CT: External Feasibility and Implications for Pulmonary Embolism AI.
Authors
Affiliations (2)
Affiliations (2)
- Department of Computer Engineering, Hallym University, Chuncheon 24252, Republic of Korea.
- Department of Emergency Medicine, Hallym University Gangnam Sacred Heart Hospital, Seoul 07441, Republic of Korea.
Abstract
CT pulmonary angiography (CTPA) is the reference test for pulmonary embolism (PE), whereas AI datasets often contain display-domain (non-DICOM) non-contrast CT (NCT). We present a three-stage pipeline: a representation-consistency gate calibrated on DICOM-verified images (Stage 1), unpaired CycleGAN translation (Stage 2), and a frequency-split composition keeping the acquired finest spatial band plus support-guided residual refinement (Stage 3). Of 208 patients, 30 were held out. The final output reduced mediastinal distribution distance (14.92 to 7.19) and contrast-gap error (11.00 to 4.25; Holm-adjusted <i>p</i> < 0.001). Distribution metrics did not differ significantly from CUT, whereas three edge-based metrics favored the proposed method in all 30 patients. On 24 external NCT scans from three manufacturers (full-pipeline feasibility analysis without paired CTPA), distribution distances decreased and edge metrics again favored the method. The gate flagged all degraded external CTPA images (blur, down-sampling, heavy compression) in every calibration resample. An external 3D CTPA-trained PE model labeled unopacified arteries as embolus on NCT; enhancement reduced this volume in normal patients (12.0 to 1.7 mL), but detection stayed at chance, as on the original NCT (AUROC 0.47; original 0.42; CTPA 0.85). The pipeline harmonizes display-domain NCT with limited structural alteration for dataset preparation rather than CTPA-based detection.