“Utilizing Machine Learning for Measuring Nuchal Translucency Thickness in First Trimester Ultrasound to Detect Chromosomal Abnormalities” رسالة ماجستير

Abstract

This study addresses the critical challenge of variability in manual Nuchal Translucency (NT) measurement by developing and evaluating a Deep Learning (DL) pipeline for automated, accurate assessment in first-trimester ultrasound images to detect chromosomal abnormalities. The research was conducted utilizing two ethically approved datasets comprising 2,425 images from pregnancies at 11-13+6 weeks gestation, sourced from Shenzhen People’s Hospital, China, and the National Hospital of Obstetrics and Gynecology, Vietnam. The study's methodology employed a two-stage DL framework. The first stage used a DenseNet121 architecture for automated image quality assessment, which successfully classified images as "Standard" or "Non-Standard" with 94% accuracy, ensuring only diagnostically reliable images were used for analysis. The second stage involved a novel DenseNet-based segmentation model to precisely delineate the NT region for measurement. This model demonstrated superior performance, achieving a Dice coefficient of 0.897 and an overall accuracy of 0.989, significantly outperforming the widely used U-Net architecture. The results confirm the pipeline's clinical efficacy. Automated NT measurements showed high agreement with expert annotations, with 73.05% of measurements deviating by less than 1mm. The integrated system achieved over 90% sensitivity and specificity for identifying high-risk cases (NT > 3mm), validating its potential as a robust decision support tool. The study's main recommendations are to integrate DL-driven quality assurance into clinical NT screening to reduce inter-operator variability, to invest in the curation of large, diverse datasets to improve model generalizability, and to foster interdisciplinary collaboration to facilitate the seamless adoption of this technology in prenatal care.