Robust landmark detection in TPUS: can AI overcome clinical image quality variability?

Image quality variability remains as a major challenge for reliable clinical deployment of artificial intelligence (AI) models in Transperineal ultrasound (TPUS). While most AI models are trained on clean, high-quality datasets, real-world clinical TPUS images often suffer from degradations caused by operator technique and acquisition parameters. This study systematically evaluates the impact of these degradations on AI models for levator ani muscle landmark localization and investigates the effectiveness of data augmentation strategies in mitigating these effects. Three widely used deep learning architectures (UNet, Spatial Configuration Network (SCN), ResNet50 encoder-decoder) were trained on good-quality set and tested on unseen realistic and synthetically degraded datasets. Moderate combined augmentation reduced mean radial error (MRE) on the good-quality test set by 17.94% (UNet), 13.29% (SCNet) and 15.56% (ResNet50) (all p<0.01), and on a realistic poor-quality set by 11.57%, 6.06% and 24.32%, respectively (all p<0.05). Shadowing, rotation and reduced brightness caused the largest drops in performance; on synthetic single degradations, moderate augmentation lowered MRE by 15.56-33.56% (UNet), 12.60-36.89% (SCNet) and 13.49-45.01% (ResNet50) versus the non-augmented baseline (all p<0.05). These findings highlight the importance of accounting for realworld image variability in AI model development for TPUS applications to ensure reliable clinical performance.