A Framework for Fetal Brain Abnormality Detection in Low-Quality Images

Abstract

The detection of fetal abnormalities through prenatal ultrasound remains a significant challenge in healthcare delivery, particularly in resource-constrained settings. Early identification of fetal brain abnormalities enables medical teams to prepare appropriate interventions, arrange for specialized delivery facilities if needed, and provide informed counseling to families regarding potential outcomes and care requirements. However, the current dependence on manual ultrasound interpretation creates workflow challenges in diagnosis, especially in regions where access to radiologists is limited. This study presents a hybrid computational approach that combines Convolutional Neural Networks (CNN), Gray Level Co-occurrence Matrix (GLCM) features, and CatBoost classification to support the detection of fetal brain anomalies in ultrasound images. Our methodology addresses specific technical challenges by incorporating bilateral filtering for preprocessing, which aims to enhance image quality in scenarios where equipment limitations or operator expertise may affect image acquisition. The system architecture integrates deep features extracted from a fine-tuned VGG16 architecture with GLCM-based texture analysis. Evaluation of the proposed methodology indicates performance metrics of 87% accuracy, 87.3% precision, 86.8% recall, and an F1-score of 86.9%. The research focuses on developing methods applicable in varying clinical environments with different levels of resource availability. The hybrid approach demonstrated higher performance metrics compared to baseline methods. This technical advancement aims to support healthcare providers in settings where radiological expertise may be limited, potentially offering a supplementary tool for ultrasound image analysis.