Dual-branch spatiotemporal-wavelet transformer conditional GAN for electrocardiographic imaging data augmentation

Electrocardiographic imaging (ECGI) reconstructs cardiac electrical activity from body-surface recordings but is limited by data scarcity and ill-posedness. We introduce a dual-branch spatiotemporal–wavelet transformer conditional GAN (STW-T cGAN) for synthesizing paired heart-surface and body-surface potential maps (HSPM–BSPM). The architecture combines Reverse Cuthill–McKee spatial reordering, axial transformers, continuous wavelet transforms (32 scales, 0.5–250 Hz), and FiLM-based adaptive fusion, with an auxiliary BSPM regression head enforcing learned cross-domain consistency. Evaluation on the ex vivo Bordeaux (primary descriptive evidence) and in vivo Auckland (exploratory corroboration) porcine datasets showed HSPM synthesis RMSE of 0.319–0.429 mV and PCC of 0.901–0.939. Auxiliary BSPM reconstruction achieved PCC 0.917–0.989. Ablation experiments demonstrated complementary branch contributions: the spatiotemporal branch improved temporal alignment while the wavelet branch enhanced spectral fidelity. FiLM fusion increased PCC by 1.7–11.4 percentage points over averaging; RCM reduced inter-electrode distance by 37%. In downstream ECGI inverse reconstruction, 1:1 augmentation decreased test-set RMSE by 6.2–14.0% across eight deep learning solvers on held-out real segmented beats (within-dataset, beat-level holdout), with attention-based architectures showing the largest improvements. Tikhonov regularization exhibited smaller gains of 3.0–3.1%. Controlled comparisons against STFT, dilated attention, and multiscale alternatives confirmed wavelet-specific benefits (5.1–28.3% RMSE reduction). These findings support STW-T cGAN as a proof-of-concept framework for paired HSPM–BSPM synthesis in data-constrained ECGI. Generalizability to human cohorts requires dedicated validation studies. • STW-T cGAN synthesizes paired HSPM–BSPM with PCC > 0.90 and SC > 0.93. • RCM reordering reduced mean epicardial inter-electrode distance by 37%. • Dual ST/CWT branches improved temporal alignment and spectral fidelity. • 1:1 augmentation reduced inverse RMSE by 6%–14% on held-out real beats (within-dataset). • FiLM fusion improved PCC by 1.7–11.4 pp over mean fusion.