A Novel Deep Feature and Oblique Decision Tree Framework for Prenatal Brain Abnormality Diagnosis

Abstract

Congenital fetal brain abnormalities affect approximately 2–3 per 1,000 pregnancies, posing significant risks to both the fetus and the mother. Early and accurate detection is crucial for timely medical intervention and improved neonatal outcomes. Conditions such as Encephalocele, Holoprosencephaly, Hydranencephaly, Intracranial Hemorrhage, Intracranial Tumor, and normal brain development require precise classification for effective prenatal care. Misdiagnosis can lead to severe complications or delayed treatment. Traditional manual analysis of fetal brain ultrasound images is time-consuming, operator-dependent, and prone to errors due to low image quality and noise. These limitations necessitate automated, reliable, and interpretable detection systems. This research work proposes an automated fetal brain abnormality detection framework using ultrasound images and advanced machine learning. Deep features are extracted from fetal brain Ultrasound images using the Convolution Attention Transformer (CoAT), which captures both local spatial details and global context. These features are initially evaluated with ensemble classifiers including eXtreme Gradient Boosting (XGBoost), Natural Gradient Boosting (NGBoost), and Histogram-based Gradient Boosting (HGBoost). To address limitations of conventional boosting models, a novel Fast Interpretable Oblique Tree (FIOT) classifier is introduced, leveraging oblique decision boundaries to model complex feature interactions while maintaining interpretability. The system categorizes fetal brain conditions into predefined classes aligned with dataset folder structures, handling class imbalance and limited training data efficiently. Its modular architecture ensures scalability, adaptability to different imaging modalities, and seamless integration with clinical workflows, providing a reliable decision-support tool for early detection of fetal brain abnormalities.