LARGE-SCALE ANNOTATION AND DEEP LEARNING FOR AUTOMATED FETAL HEAD BIOMETRIC IDENTIFICATION IN ULTRASOUND IMAGES

Abstract

This thesis encompasses a series of methodologies in ultrasound imaging analysis, aimed at enhancing fetal health monitoring during gestation. It commences with an end-to-end framework for segmenting, measuring, and estimating fetal gestational age and weight using two-dimensional ultrasound images. This framework, integrating eight segmentation architectures and a weighted voting method, forms an ensemble transfer learning model. It achieves a segmentation accuracy of 98.53% mean intersection over union (mIoU) and measurement accuracy with a 1.87 mm mean absolute difference (MAD). Additionally, it predicts the week of gestational age with a mean square error (MSE) of 0.03% and the estimated fetal weight (EFW) with an MSE of 0.05%. The research progresses to address speckle noise in ultrasound images. A composite image technique is identified as effective in improving segmentation performance, showing an mIoU of 0.96893%, mean pixel accuracy (mPA) of 0.97831%, and an average peak signal-to-noise ratio (PSNR) of 53.034 dB. Further, a method for converting pixel values to millimeters in fetal ultrasound images is introduced, facilitating more accurate fetal measurements. This method, applied to an augmented dataset of 2835 fetal head images, proves critical for AI applications in automated fetal measurements. A deep learning-based solution for automating the labeling process in AI applications is also presented. The Xception model demonstrates a high R-squared value of 0.8535 and a mean squared error (MSE) of 0.00028 when predicting pixel size. The creation of a comprehensively annotated ultrasound fetal head dataset of 3,832 highresolution images constitutes another part of this research. This dataset, validated for quality and reliability, provides a vital resource for developing computer vision algorithms in prenatal diagnostics. Finally, the thesis introduces FetSAM (Fetal Segment Anything Model), a deep learningbased segmentation model for fetal head biometrics in ultrasound imagery. FetSAM achieves a Dice Similarity Coefficient (DSC) of 0.90117, Hausdorff Distance (HD) of 1.86484, and Average Surface Distance (ASD) of 0.46645, indicating high precision in segmenting fetal structures. This thesis contributes to the field of medical imaging and artificial intelligence, particularly in prenatal diagnostics. The developed methodologies and models enhance the accuracy and efficiency of fetal health monitoring, supporting advancements and clinical applications in this domain.

Date of Award	2023
Original language	American English
Awarding Institution	HBKU College of Science and Engineering