UBC Theses and Dissertations

UBC Theses Logo

UBC Theses and Dissertations

Automated analysis of the placenta in ultrasound Hu, Ricky


The placenta is an organ that serves as an interface for macromolecule exchange between a mother and fetus. Revealing symptoms of placenta disease are usually presented later on in the pregnancy with limited treatment options. Access to specialists of placental disease is also limited, particularly in regions distant from urban centres. There is a need for an accessible method to screen patients for risk of placenta disease early. Ultrasound provides a non-invasive imaging modality capable of visualizing the placenta commonly used in obstetric clinics. However, ultrasound images contain unique artifacts that are difficult to interpret with visual inspection. This thesis presents a pipeline of three ultrasound processing methods developed to aid placenta ultrasound analysis. The first is a method to detect ultrasound acoustic shadow artifacts that obscure anatomy. Radiofrequency signals and pixel entropy were analyzed to identify acoustic shadow in images. A clinical study was performed to obtain ultrasound scans from 37 subjects to evaluate performance. A Dice coefficient of 0.90±0.07 for radiofrequency-based and 0.87±0.08 for pixel entropy-based techniques was achieved when compared to manual shadow detection. The second is a method to segment the placenta in images preprocessed by shadow detection using a convolutional neural network. Performance was evaluated on data from 1364 fetal ultrasound images from 247 patients. A Dice coefficient of 0.92±0.04 was achieved when compared to manual segmentation. The third is a method to classify placenta appearance as either normal or abnormal. Images were preprocessed with the first two methods to provide a placenta-only image. A residual convolutional neural network was then used to classify the placenta appaearance. Performance was evaluated on 7831 fetal ultrasound images from 367 patients. Placenta classification achieved a sensitivity of 0.91 and a specificity of 0.87 when compared to classification by physicians. The methods demonstrate the capability of ultrasound physics analysis and machine learning methods in processing placenta ultrasound images. The results show the potential for developing a tool in the future to assist physicians in analyzing the placenta to screen for disease.

Item Media

Item Citations and Data


Attribution-NonCommercial-NoDerivatives 4.0 International