Year: 2026
Type: article
Abstract: Prenatal ultrasound is the primary imaging modality for monitoring fetal growth and screening for congenital anomalies, yet its interpretation depends on operator experience, image quality and the availability of specialists. This paper presents the design and interim progress of a deep-learning framework for AI-assisted, multi-disease fetal anomaly detection from prenatal ultrasound images. The framework combines patient-wise data splitting, ultrasound-specific preprocessing, a baseline convolutional neural network (CNN) and transfer learning with ImageNet-pretrained ResNet50, DenseNet and EfficientNet models. It is planned to be extended to multi-disease classification (brain, heart, spine, kidney and other anomalies), Grad-CAM explanations and a Flask-based web interface. A review of six core studies identifies six research gaps: limited multi-disease coverage, small and imbalanced datasets, a weak link between plane recognition and anomaly diagnosis, limited external validation, weak early-stage detection, and limited interpretability. Progress recorded up to October 2026 (approximately 35% of the project) comprises dataset acquisition, cleaning, preprocessing, patient-wise train/validation/test splitting, a baseline CNN in TensorFlow/Keras, and initial fine-tuning experiments with three pretrained networks. Early observations are overfitting, bias toward majority classes and run-to-run variability. Multi-disease classification, full evaluation, Grad-CAM, the web application and testing remain pending, and no performance figures are claimed. The system is a research prototype and not a diagnostic device
Institution: Sandip Institute of Technology & Research Centre
Language: English
Field: Medicine
Domain: Health Sciences
Open Access status: diamond
OpenAlex ID: W7224497619