Deep Learning Based Detection of Poor Probe Contact in Liver Ultrasound Images for Enhanced Shear Wave Elastography Acquisition
ePoster
Authors: Abder-Rahman Ali, Massachusetts General Hospital/Harvard Medical School Anthony Samir, Massachusettes General Hospital/Harvard Medical School
Our abstract introduces a novel approach to an underexplored issue in liver ultrasound imaging - the detection of poor probe contact. We present SimCLR+ENet, a unique two-stage deep learning model, trained on a substantial dataset of unannotated images. This large unannotated dataset is the cornerstone of our approach, enabling us to leverage the power of SimCLR to extract meaningful feature representations and label a feasible number of annotated images (i.e., for ENet), making the task manageable and efficient for domain experts involved in the annotation process. Our model demonstrates an overall accuracy of 65.4% in detecting poor probe contact, with a sensitivity and specificity of 49.1% and 82.1% respectively. By automating the detection of poor probe contact, we can enhance the quality of liver ultrasound images and the accuracy of liver stiffness measurements during shear wave elastography (SWE) acquisition, potentially revolutionizing liver fibrosis staging. Future work will address the zooming factor during training and the need for a larger dataset of annotated poor probe masks to improve the performance of the proposed model. Attendees will gain insights into the transformative potential of deep learning in liver ultrasound imaging and the exciting possibilities for further refinement and improvement of our model.