An interpretable deep stacked autoencoder architecture for rapid convergence and robust genetic disorder detection from fetal ultrasound images

Early and reliable detection of fetal genetic disorders (FGD) using prenatal ultrasound (US) imaging is crucial for timely treatment intervention and improved postpartum outcomes. US images are inherently affected by speckle noise, reduced contrast, and subtle anatomical variations, which impede automated diagnosis. This study presents the Stacked Autoencoder Explainable Artificial Intelligence Network (SAE-XAI-Net) for FGD Classification, a deep learning (DL) framework aimed at accurate and interpretable classification of FGD. The proposed architecture employs contrast-limited adaptive histogram equalization (CLAHE) to enhance US images, thereafter learning hierarchical features using a stacked convolutional autoencoder. The encoded latent representations are then refined using a supervised classification module, facilitating feature differentiation and expediting the process. Gradient-weighted Class Activation Mapping (Grad-CAM) is used to elucidate model predictions for physicians by highlighting physically significant regions that influence them. They evaluated our methodology using a multi-class prenatal US dataset, including both normal cases and those with chromosomal disorders. The proposed SAE-XAI-Net achieves a prediction accuracy of 96.4%, surpassing the performance of conventional convolutional neural network (CNN) baselines. The model demonstrated commendable accuracy, recall, F1-score, and area under the receiver operating characteristic (ROC) curve (area under the curve (AUC)), indicating its robustness and applicability across many scenarios. The framework’s rapid convergence was particularly noteworthy; it achieved near-optimal performance during the first training epochs while maintaining constant validation accuracy. The explainability investigation verified that the model consistently focused on clinically relevant fetal characteristics throughout classification, hence validating its suitability for medical use. SAE-XAI-Net is a dependable, comprehensible, and rapid method for using computers to test for genetic diseases in pregnant women. It has significant potential for use in practical obstetric diagnostic procedures.