Automatic sonographer skills assessment

Thesis

Sonography is widely recognised as a difficult task to learn. Skills for routine fetal ultrasound screening are traditionally learned and assessed under supervision and with feedback from senior sonographers. This thesis aims to characterise sonographer skill differences in routine fetal ultrasound scanning. There is no universally agreed method for differentiating ultrasound skill levels and consequently a gold standard ground-truth grading for sonographers or their scans does not exist. Therefore, in this work we explore different aspects of sonographer skills and build deep learning-based models to describe these skills.

In our first contribution, we investigate the differentiation of sonographer skills between two skill level groups using probe motion tracking data. Skill level is defined by the number of years of scanning. Operators with two or more years of experience in fetal anomaly screening ultrasound scanning are defined as “expert” level, and those with less than two years of experience as “newly qualified”. We propose a deep learning-based framework to model the probe motion data during routine fetal ultrasound scanning and predict the skill level from the input motion signals.

For our second contribution, we consider sonographer’s skill at performing two biometry (measurement) tasks in fetal ultrasound examination; measuring the head circumference and measuring the abdomen circumference on ultrasound video frames. Firstly, for each task, a frame classification model is trained to produce a binary prediction to indicate whether the input frame is a biometry plane. Secondly, we propose to estimate operator skill levels by evaluating task models performance on the acquired data and develop models to predict skill levels based on ultrasound video and synchronised probe motion data. Two criteria related to frame classification task performance were proposed for training the skill predictors.

In a third contribution, we extend the second contribution by considering the co-dependence between the task model and the skill model, utilising the task of anatomical landmark segmentation for finding the head circumference measurement plane. In this work, we propose a bi-level optimisation model trained on ultrasound video frames, including an upper-level skill predictor and a lower-level anatomical landmark segmentation task predictor, which are optimised jointly by refining the two networks simultaneously. Once trained, the upper-level skill predictor network can be used independently without the task predictor network to give a skill prediction from input ultrasound frames.

The models proposed in the thesis have been tested on retrospective data and will hopefully inspire future work to develop deep learning-based models to assist in the training of fetal ultrasound sonographers and provide automatic informative skills assessment.

Authors: Wang, Y

• DOI: 10.5287/ora-nbjbvqwpo
• Type of award: DPhil
• Level of award: Doctoral
• Awarding institution: University of Oxford
• Language: English