Detecting Fetal Brain Abnormalities from Ultrasound Images by Deep Learning Methods

The earliest and most accurate outcome of fetal brain abnormality can lead to timely medical intervention and the substantial improvement of neonatal outcomes. A method using deep learning to automatically analyze ultrasound images for finding structural defects in the fetal brain. The system fuses multiple neural network models to grasp spatial and contextual information, which is most likely to be relevant to 2D ultrasonography. A series of well-thought-through image processing methods enables the input data to be of high quality and consistent, which in turn allows the deep models to learn better. It has been also added a new classification layer to their network, which enables the system to achieve an accuracy of 96.3% and also learns to recognize normal and abnormal patterns depending on the very small shape differences that human experts have often missed, even during screening. Besides, it also uses a fusion strategy, which allows combining the best parts of several deep learning methods for a more accurate detection of the abnormality. The approach introduces a fully automated and scalable solution that can be used in clinical workflow, as it not only helps clinicians in making decisions but also decreases observer variability and allows for early management of fetal neurological disorders.