StackFBAs: Detection of fetal brain abnormalities using CNN with stacking strategy from MRI images

Abstract

Predicting fetal brain abnormalities (FBAs) is an urgent global problem, as nearly three of every thousand women are pregnant with neurological abnormalities. Therefore, early detection of FBAs using deep learning (DL) can help to enhance the planning and quality of diagnosis and treatment for pregnant women. Most of the research papers focused on brain abnormalities of newborns and premature infants, but fewer studies concentrated on fetuses. This study proposed a deep learning-CNN-based framework named StackFBAs that utilized the stacking strategy to classify fetus brain abnormalities more accurately using MRI images at an early stage. We considered the Greedy-based Neural architecture search (NAS) method to identify the best CNN architectures to solve this problem utilizing brain MRI images. A total of 94 CNN architectures were generated from the NAS method, and the best 5 CNN models were selected to build the baseline models. Subsequently, the probabilistic scores of these baseline models were combined to construct the final meta-model (KNN) utilizing the stacking strategy. The experimental results demonstrated that StackFBAs outperform pre-trained CNN Models (e.g., VGG16, VGG19, ResNet50, DenseNet121, and ResNet152) with transfer learning (TL) and existing models with the 5-fold cross-validation tests. StackFBAs achieved an overall accuracy of 80%, an F1-score of 78%, 76% sensitivity, and a specificity of 78%. Moreover, we employed the federated learning technique that protects sensitive fetal MRI data, combines results, and finds common patterns from many users, making the model more robust for the privacy and security of user-sensitive data. We believe that our novel framework could be used as a helpful tool for detecting brain abnormalities at an early stage.