Fetal Organ Anomaly Identification, Segmentation, and Optimized Data Augmentation

Early detection of fetal diseases enables clinicians to provide early interventions and treatment. In this thesis, several contemporary techniques are proposed and combined to help detect and identify affected organs. We propose a novel approach to fetal Magnetic Resonance Imaging (MRI) affected-organ classification that consists of whole fetal body segmentation, which is both integrated with Data Augmentation (DA) optimization for optimal image transformations. We first present a novel segmentation architecture using attention mechanisms and Squeeze-and Excitation (SE) modules to emphasize localization of contextual information that is relevant in biomedical segmentation. We use Naïve Inception (NI) and SE modules to create an efficient classification algorithm. Finally, we conducted detailed ablation studies to show that the proposed combined segmentation-classification-augmentation pipeline is superior in accuracy and performance to the individual components. Our results show that we can improve classification from 85.32% to 88.91% when identifying the affected organs in fetal MRI.