Adversarial domain adaptation for false-positive reduction in hospital-to-robotic echocardiographic screening under limited target-domain data.
Authors
Affiliations (2)
Affiliations (2)
- Department of Modern Mechanical Engineering, Graduate School of Creative Science and Engineering, Waseda University, Tokyo, Japan.
- Iwata Laboratory, Department of Modern Mechanical Engineering, Graduate School of Creative Science and Engineering, Waseda University, Tokyo, Japan.
Abstract
Hospital-trained deep learning models for echocardiographic wall motion abnormality detection achieve strong performance under standard clinical acquisition conditions but often degrade substantially when deployed on autonomous robotic acquisition systems. Most existing models are developed and validated under controlled clinical acquisition conditions, limiting their robustness when deployed in emerging autonomous imaging systems. A hospital-trained wall motion abnormality detection model trained on 10,000 hospital echocardiograms achieved 92.3% accuracy (94.0% sensitivity, 90.6% specificity) when applied to an independent hospital test data set of 500 studies. However, direct deployment on 150 robotic echocardiograms from 15 healthy subjects reduced the normal-case specificity to 65.7%, representing a 24.9 percentage-point drop compared to the performance when applied to the hospital-acquired data. No statistically significant difference in SNR was observed between domains (p = 0.31; d = 0.12), with similarly small standardized differences in contrast and sharpness, although formal equivalence was not established. Acquisition-geometry measures differed descriptively between domains, suggesting that geometry-associated factors may contribute to the observed domain shift. We propose an adversarial domain adaptation framework built upon a 3D ResNet-18 backbone with a gradient-reversal domain discriminator for aligning hospital and robotic feature representations while retaining the hospital-trained classification objective. This design enables adaptation to the robotic domain without requiring labeled robot-acquired pathological cases, which are particularly difficult to obtain during early-stage robotic system development. Under leave-one-subject-out validation across robotic subjects, Mixed-DANN improved specificity to 86.1% ± 4.3%, while Normal-only DANN improved specificity to 89.3% ± 4.8% (p < 0.001 versus direct transfer), recovering 94.8% of the hospital-to-robot specificity gap. Because the robotic cohort comprises only healthy subjects, evaluation is appropriately restricted to specificity and false-positive rate; sensitivity evaluation requires future acquisition of robot-acquired pathological cases. These findings suggest that acquisition-associated domain shift substantially impacts cross-platform deployment. The proposed adversarial domain adaptation framework provides a potential pathway for adapting hospital-trained cardiac AI models to autonomous robotic ultrasound under conditions of limited target-domain data, although validation using robot-acquired pathological cases remains necessary before clinical translation.