Artificial intelligence for intracardiac echocardiography interpretation and guidance during atrial fibrillation ablation
Year: 2026
Type: article
Abstract: Background Intracardiac echocardiography (ICE) has becoming indispensable for ensuring safety and efficacy during atrial fibrillation (AF) ablation. However, ICE interpretation remains highly observer-dependent, presenting a steep learning curve in rapidly distinguishing complex, dynamic cardiac structures. Objectives This study aimed to develop and validate a multi-task deep learning system (AI-ICE) for real-time, automated interpretation of cardiac structures and ultrasound views during AF ablation. Methods The interpretation task is defined as two sub-tasks, one for detecting 12 anatomical structures and the other for classifying 5 standard views. A total of 373,091 ICE images from 350 patients undergoing AF ablation across three centers were collected. The dataset was split into training/internal validation cohorts (two centers) and an independent external testing cohort (one center, 35,335 frames). The AI-ICE model combined a ResUnet backbone with an LSTM module to capture spatiotemporal dependencies across consecutive frames, reflecting the streaming nature of real-time clinical ICE imaging. The two tasks were formulated as a multi-task coupled framework with shared representations, exploiting the intrinsic interdependence between the imaging view and the visible anatomy. Results For the internal validation, the average AUC for classifying ultrasound views and detecting structures reached 0.971 (95% CI: 0.947–0.992) and 0.974 (95% CI: 0.968–0.980). For the external validation, the AI-ICE system achieved an AUC of 0.984 (95% CI: 0.974–0.993) for ultrasound view classification and 0.955 (95% CI: 0.948–0.962) for anatomical structure objection. Compared with three cardiologist experts, it achieved higher accuracy (AUC: 0.941 vs. 0.916) and superior inter-rater agreement (Kappa coefficient: 0.791 vs. 0.716). Additionally, the system demonstrated high efficiency, with an inference time of 12 milliseconds per ICE image. Conclusions The multi-task AI-ICE system enables real-time, operator-independent ICE interpretation during AF ablation, establishing a foundation for automated intra-operative guidance.
Language: English
Subfield: Health Informatics
Field: Medicine
Domain: Health Sciences
Open Access status: gold
Estimated APC paid: $3,801
OpenAlex ID: W7226029052
