Fetus Abnormality Detection and Classification Using GradCAM & EfficientNet-B0 with Deep Learning Segmentation

Abstract

Prenatal diagnosis of fetal structural abnormalities via ultrasound is a critical yet challenging task, heavily dependent on sonographer expertise and subjective interpretation. This paper presents a novel deep learning framework for automated fetal abnormality detection and multi-class classification using EfficientNet-B0 integrated with Gradient-weighted Class Activation Mapping (GradCAM) for explainable visual reasoning and U-Net-based segmentation for anatomical region isolation. The proposed system processes fetal ultrasound images to detect and classify anomalies across six categories: neural tube defects, cardiac malformations, abdominal wall defects, skeletal dysplasia, facial clefts, and chromosomal markers. EfficientNet-B0 leverages compound scaling to achieve optimal accuracy with minimal parameters (5.3M), outperforming heavier architectures such as VGG-16 and ResNet-50 while maintaining real-time inference speed. GradCAM visualization overlays highlight diagnostically relevant regions, providing interpretable outputs for clinical validation. The model was trained on a curated dataset of 18,500 annotated fetal ultrasound images spanning 12 gestational weeks (18–30 weeks), achieving an overall classification accuracy of 96.4%, precision of 95.8%, recall of 94.9%, and F1-score of 0.953, with a mean AUC-ROC of 0.971. Segmentation performance yielded a mean Intersection-over-Union (mIoU) of 0.887. Comparative evaluation demonstrates superiority over existing CNN-based approaches. The system provides an end-to-end diagnostic aid suitable for integration into clinical ultrasound workflows, reducing diagnostic time by 68% and supporting early intervention planning.