Construction and clinical validation of a fetal brain magnetic resonance imaging-prediction model based on multimodal AI fusion algorithm.

OBJECTIVE

To develop a multimodal artificial intelligence (AI) fusion model predicting abnormal fetal brain development from magnetic resonance imaging (MRI).

METHODS

Using fetal brain MRI data and clinical indicators from pregnant women (from January 2021 to December 2023), who were split 7:3 into training and validation sets. In the training set, key predictors were identified via univariate analysis and multivariate logistic regression, including both clinical indicators and continuous MRI biometric parameters. Three multimodal AI fusion models，including Convolutional Neural Network-Recurrent Neural Network (CNN-RNN) model, attention mechanism-based model, and feature concatenation model were developed. Performance was assessed by accuracy, precision, recall, F1-score, and the area under the receiver operating characteristic curve (AUC).

RESULTS

Among the total 806 participants, 108 cases (19.15 %) had fetal brain abnormalities in the training set (n = 564), 45 cases (18.59 %) in the validation set (n = 242). Multivariate logistic regression analysis showed that gestational age, gestational diabetes mellitus, alpha-fetoprotein, lateral ventricular width, and sulcation development score were independent risk factors for fetal brain abnormalities. The attention mechanism fusion model achieved the highest AUC in both the training set (0.876) and the validation set (0.869), significantly outperforming the CNN-RNN fusion model (AUC in training set: 0.776; AUC in validation set: 0.718) and the feature concatenation fusion model (AUC in training set: 0.754; AUC in validation set: 0.720).

CONCLUSION

The multimodal AI fusion model, particularly using attention mechanisms, effectively identifies high-risk fetal brain abnormalities, offering potential for early clinical intervention and improved prenatal counseling to enhance detection and prognosis of neurological disorders.