Early detection of fetal brain anomalies is critical for ensuring appropriate medical care. This paper presents an improved deep learning approach integrating image super-resolution with advanced classification techniques. Unlike prior work, we introduce a custom fine-tuned deep learning pipeline that enhances low-resolution ultrasound images before classification. The novel proposed architecture incorporates a modified Enhanced SRGAN for super-resolution and an optimized CNN classifier integrating VGG16, ResNet, and DenseNet features. A dataset of 4,000 grayscale ultrasound images (512×512 pixels) was collected and categorized into four classes: normal, cerebellum anomalies, thalamic anomalies, and ventricular anomalies. To address class imbalance (1039 normal vs. 2961 abnormal), oversampling, augmentation, and class-weighted loss functions were applied. Unlike previous studies, we provide a comprehensive performance analysis using accuracy (92.94%), precision, recall, F1-score, and confusion matrices, demonstrating the impact of super-resolution on classification accuracy. This research significantly improves fetal brain anomaly detection and establishes a robust deep learning pipeline for clinical applications.
A Sun, study studied this question.