Ultrasound Scans are widely used to visualize the development of fetuses during pregnancies where Fetal Growth, Gestational Age, and Fetal Anomalies are observed. Fetal Anomalies of the cranial region can be detected by an examiner most of the time. However, due to ultrasound feeds being noisy and dimly illuminated, examiners often need to conduct careful observations of fetuses with limited technical support.
The aim of this project is to combine existing methodologies and automatically detect fetal anomalies in the cranial region using Machine Learning and Medical Imaging. Such problems can be solved using the four-phase methodology: Preprocessing, Segmentation, Feature Extraction, and Classification.
The preprocessing stages included annotating fetal heads and artificially creating anomalous cases to balance the dataset. UNet, a deep learning model, was used for the semantic segmentation of the fetal head. Feature Extraction was done manually after using ellipse fitting techniques to measure the head. Finally, classification was done using two variants of the Logistic Regression model: Simple Logistic Regression and Multinomial Logistic Regression.
The validation accuracy of UNet was 96.34% and the accuracy of the classifier was 82.03%. In summation, the project seemed to propel in cases where anomalies were on the extreme ends (high risk or normal) but showed signs of misclassification in the Low-Risk cases. Further improvements can be made by balancing the number of images in classes of the dataset and rigorously preprocessing noisy regions in the existing images.