Medical image analysis for simplified ultrasound protocols

Thesis

Ultrasound is an imaging tool used in obstetrics to identify high-risk pregnancies. However, ultrasound (US) requires a trained operator, who guides a transducer in response to real-time interpretation of video content. In low- and middle-income countries (LMICs), there is a shortage of trained sonographers. In this thesis, we address this key challenge by combining simple US video sweeps with computational algorithms to provide clinical benefit. The sweeps can be taken by an US novice. First, we design an algorithm that automatically creates an assistive video overlay from a simple video sweep. The overlay assists interpretation of US video to assess placenta location. We describe the design and evaluation of a deep learning-based automatic segmentation model and a statistical data visualisation of 2-D placenta shapes. The data visualisation reveals the spectrum of placenta shapes in this problem space. A probabilistic graphical model is used to improve segmentations with regards to the highly variable placenta shape. From the automatic segmentations, image guidance is created, translating the clinical criteria into assistive visual information. Second, we explore analysis of multiple video sweeps using graphs. A three-node graph models three video sweeps, where the nodes encode binary sequences representing the fetal head frame-level detection across all video frames in a sweep. To better characterise the sweeps, we perform a statistical analysis of large-scale manual annotations of video sweeps in our dataset. This reveals common patterns of frame-level anatomy occurrence for different video sweep trajectories. Particular insight is gained for patterns that correspond to fetal pose. In this regard, we build a graph convolutional network to automatically classify fetal presentation, using graphs that combine complementary video sweep information relating to fetal pose. Finally, we demonstrate the feasibility of placenta 3-D reconstruction using multiple video sweeps. We pose this challenging problem as spatio-temporal alignment of US video. We first temporally align video sweeps to represent video content at the same temporal scale. Then, we use affine transformations to spatially align images in temporally aligned video. The results in this chapter are exciting as they show the feasibility of placenta 3-D reconstruction in a simple US sweep system.

Authors: Gleed, AD

• DOI: 10.5287/ora-4jvegrkdd
• Type of award: DPhil
• Level of award: Doctoral
• Awarding institution: University of Oxford
• Language: English
• Keywords: Placenta; Simplified ultrasound protocols; Medical image analysis; Video analysis; Obstetrics; Ultrasound; Fetal imaging; Global health; Machine learning; Deep learning; Ultrasound video sweeps
• Subjects: Deep learning (Machine learning); Biomedical engineering; Machine learning; Engineering; Image segmentation; Image processing