AI learning for pediatric right ventricular assessment: development and validation across multiple centers

Figures

• 

  Fig. 1. RV function assessment powered by AI learning of multi-center pediatric echocardiograms. A Performance analysis, on U.S. single-center data (training and validation sets), of the Dice similarity coefficient for the overall, end-systolic, and end-diastolic tracing. B The predicted RV FAC compared with the reported FAC for the U.S. single-center data (training and validation sets). MAE: Mean Absolute Error. RMSE: Root Mean Square Error. C Receiver-operating characteristic curves for the RV function assessment with reduced RV FAC (<35%) for the U.S. single-center data (training and validation sets). D The predicted RV FAC compared with the reported FAC for the Asia multi-center validation data. E Receiver-operating characteristic curves for the RV function assessment with reduced RV FAC (<35%) for the Asia multi-center validation data.

• 

  Fig. 2. RV-based classification model performance: CSN#1 modeling detection of PH from normal conditions and CSN#2 modeling differentiating PH from TOF. A U.S. single-center testing results; B Asia multi-center testing results; C validation performance measures.

• 

  Fig. 3. Consort diagram of patient enrolment.

• 

  Fig. 4. Multi-center RV AI learning workflow to predict FAC and allow automated assessment of RV function.

• 

  Fig. 5. Multi-center RV-based AI learning workflow for PH classification. The workflow begins with the construction of a training cohort from a U.S. single center, used to train a U²-Net model for RV segmentation. Segmentation results enable beat-by-beat analysis of echocardiographic videos, which structures the input for CSN models. The CSN#1 and CSN#2 models are then trained to perform classification: CSN#1 distinguishes between normal and PH cases, while CSN#2 distinguishes between PH and TOF. The trained models are validated on the remaining U.S. data and tested on a multi-center Asian cohort. Solid arrows indicate the training process; dashed arrows indicate validation and testing pathways. PH pulmonary hypertension, TOF Tetralogy of Fallot, CSN channel-separated convolutional network.