Application of deep learning models to transesophageal echocardiography images for percutaneous left atrial appendage closure
R Tateishi, T Umemoto, M Shimizu, Y Matsuda, I Kawamura, K Goto, H Shimada, N Kato, M Suzuki, S Miyazaki, T SasanoAbstract
Background/Introduction
Percutaneous left atrial appendage closure (LAAC) with WATCHMAN FLX™ offers stroke prevention comparable to OAC and often enables discontinuation. Nonetheless, procedures may be aborted because of unfavourable appendage size or morphology, and complex anatomies can prolong manipulation or necessitate multiple devices, increasing burden and cost. Convolutional neural networks (CNNs), a deep-learning method, have shown strong image-recognition performance in cardiology.A deep-learning, image-centric strategy may provide a more objective and reproducible basis for device size selection.
Purpose
To develop and validate a CNN using transoesophageal echocardiography (TEE) to assist WATCHMAN FLX™ size selection, and to apply Gradient-weighted Class Activation Mapping (Grad-CAM) to visualise model attention and identify anatomy influencing predictions.
Methods
We retrospectively analysed adults who underwent LAAC with WATCHMAN FLX™/FLX Pro™ at two centres between August 2021 and December 2024. Procedures were performed with continuous TEE guidance; device choice was based on TEE and angiographic findings, and implantation proceeded only when PASS (Position, Anchor, Size, Seal) criteria were met. Intraprocedural TEE images acquired left atrial appendage (LAA) views at 0°, 45°, 90°, and 135° once left atrial pressure was ≥10 mmHg. From each sequence, the most informative still frames were assembled into four-view composite images (up to 10 per patient; Figure 1), and images of inadequate quality were excluded. For CNN development, images from both centres were randomly split 4:1 into training and validation sets. We compared VGG-19, EfficientNet-B2, and EfficientNetV2-S; the architecture with the highest validation accuracy was selected as the final model.
Results
Seventy-four patients underwent attempted LAAC (mean age 75.1 years; 78.4% male). Median CHA2DS2-VASc was 4.5 and HAS-BLED 3; 33.8% had paroxysmal and 66.2% persistent AF. Echocardiography showed preserved LVEF (60.4%) and moderate left atrial enlargement (mean diameter 44.7 mm). WATCHMAN FLX™ was selected initially in 56 patients and FLX Pro™ in 18; three cases (4.1%) were unsuccessful, with no centre-level difference. The success rate with the initially selected device was 81.1%; 14 patients required multiple devices. A total of 590 images were incorporated into the CNN dataset. Among architectures, EfficientNet-B2 achieved the best validation performance (training accuracy [0.8475], validation accuracy [0.7542]). Grad-CAM predominantly highlighted the 0° view around the LAA, indicating attention to features pertinent to WATCHMAN sizing (Figure 2).
Conclusion
The CNN model showed performance in line with conventional TEE-guided sizing (first-device success rate 81.1%). Explainable AI via Grad-CAM localises salient LAA regions, highlighting features relevant to device sizing and may support standardised, AI-assisted pre-procedural planning.