Real-Time Ultrasound Brightness Correction Under Misalignment and Limited Supervision

Brightness mode ultrasound is strongly affected by operator-controlled gain and time-gain compensation (TGC). Incorrect settings can obscure anatomical boundaries, reduce contrast, and impair interpretability, particularly in point-of-care and educational use. This study evaluates a conservative artificial intelligence approach for real-time ultrasound brightness correction under limited supervision and residual misalignment. Instead of synthesizing a new image freely, the corrected output is constrained to be the acquired frame multiplied by a predicted gain map. Two generative adversarial network-based architectures were compared: RetinexNet, a higher-capacity U-Net-style model with temporal attention, and GeoGainNet, a lightweight geometry-aware model that factorizes log-gain into global, depth-profile, and residual components. The models were evaluated with paired multi-split experiments on a real paired carotid ultrasound benchmark and a synthetic EchoNet-derived curved-probe benchmark. Evaluation combined paired image quality, TGC-profile behavior, gain-map fidelity, anatomy preservation, temporal consistency, expert reader assessment, and CPU/GPU runtime.On the synthetic EchoNet benchmark, both models improved 18 of 19 held-out test metrics, and all positive improvements remained significant after Benjamini-Hochberg false discovery rate correction. RetinexNet showed the stronger synthetic quality profile, outperforming GeoGainNet on 16 of 19 test metrics. On the real paired benchmark, RetinexNet improved 16 of 19 test metrics and GeoGainNet improved 18 of 19, with 10 and 8 false-discovery-rate-significant improvements, respectively. The direct real-data model comparison was balanced, whereas runtime strongly favored GeoGainNet: median GPU throughput was 294.6 frames/s versus 4.9 frames/s for RetinexNet, and median CPU throughput was 18.70 frames/s versus 0.24 frames/s. A single-expert reader evaluation on 15 real cases gave similar mean scores for both models. These findings support multiplicative gain-map learning as an interpretable basis for ultrasound brightness correction, while highlighting the need for larger real-data and multi-reader validation.