[Industry] Anatomy highlighting in lung ultrasound Images via deep learning

James Tapp


Supervised by Oktay Karakus; Moderated by James Osborne

(DSA-23) Ultrasound is a non-invasive and safe (non-ionising) imaging medium which may be used to diagnose and monitor many conditions, including the assessment of lung pathologies. At Intelligent Ultrasound we enhance the effectiveness of Ultrasound-based diagnoses through the use of AI and in this project, we’d like your help!

We have a large (though only partially labelled) dataset of lung ultrasound images, containing tens of thousands of images obtained from hundreds of patients. The project will develop a system for highlighting relevant features of the image (for example, pleura, consolidations, A-lines, and B-lines) during live scanning, using Deep Learning models. Such a system, if sufficiently effective can be put into clinical use to assist newly trained clinicians especially in interpreting ultrasound, reducing the risk of misdiagnosis.

The models will need to be capable of highlighting anatomical features very accurately while also running at more than 10 FPS on modest hardware. A key differentiator in generating the best performance from the available data may be the use of semi-supervised learning techniques to extract value from the unlabelled portion. In addition to providing data, we will make our GPU cluster available for model training.

The project will be a challenging but interesting opportunity to apply Deep Learning techniques to a large real-world data set, with the chance to create a system that would have real clinical utility.

The deliverables for the project would be:

1) Train a Deep Learning model on our lung ultrasound dataset which can accurately produce anatomical highlighting while also running at a sufficiently high frame rate. Optionally explore potential improvements that might stem from applying semi-supervised learning approaches.

2) Evaluate the effectiveness of the model produced and present your findings to the R&D team at Intelligent Ultrasound.

Final Report (09/10/2023) [Zip Archive]

Publication Form