Assessment of obstetric ultrasound images using machine learning

Ultrasound-based fetal biometry is used to derive important clinical information for identifying IUGR (intra-uterine growth restriction) and managing risk in pregnancy. Accurate and reproducible biometric measurement relies heavily on a good standard image plane. However, qualitative visual assessment, which includes the visual identification of certain anatomical landmarks in the image is prone to inter- and intra-reviewer variability and is also time-consuming to perform. Automated anatomical structure detection is the first step towards the development of a fast and reproducible quality assessment of fetal biometry images. This thesis deals specifically with abdominal scans in the development and evaluation of methods to automatically detect the stomach and the umbilical vein within them.First, an original method for detecting the stomach and the umbilical vein in fetal abdominal scans was developed using a machine learning framework. A classifier solution was designed with AdaBoost learning algorithm with Haar features extracted from the intensity image. The performance of the new method was compared on different clinically relevant gestational age groups.Speckle and the low contrast nature of ultrasound images motivated the idea of introducing features extracted from local phase images. Local phase is contrast invariant and has proven to be useful in other ultrasound image analysis application compared with intensity. Nevertheless, it has never been implemented in a machine learning environment before. In our second experiment, local phase features were proven to have higher discriminative power than intensity features which enabled them to be selected as the first weak classifiers with large classifier weight.Third, a novel approach to improving the speed of the detection was developed using a global feature symmetry map based on local phase to select the candidate locations for the stomach and the umbilical vein. It was coupled with a local intensity-based classifier to form a “hybrid” detector. A nine-fold increase in the average computational speed was recorded along with higher accuracy in the detection of both the anatomical structures.Quantitative and qualitative evaluations of all the algorithms were presented using 2384 fetal abdominal images retrieved from the image database study of the Oxford Ultrasound Quality Control Unit of the INTERGROWTH-21st project.Finally, the “hybrid” detection method was evaluated in two potential application scenarios. The first application was clinical scoring in which both the computer algorithm and four experts were asked to record presence or absence of the stomach and the umbilical vein in 400 ultrasound images. The computer-experts agreement was found to be comparable with the inter-expert agreement. The second application concerned selecting the standard image plane from 3D abdominal ultrasound volume. The algorithm was successful in selecting 93.36% of the images plane defined by the expert in 30 ultrasound volumes.