MIDL Vision 2020

Luyao Shi has a lot to be proud of. As if having a paper accepted at MIDL wasn’t enough, he has just completed his PhD at Yale University in the Department of Biomedical Engineering. The work he is presenting here is about an automatic pulmonary embolism diagnosis framework using deep learning. He worked on it during his internship at IBM Research last summer and hopes to return there after graduation. He speaks to us ahead of his oral and poster session today. Pulmonary embolism (PE) is a life- threatening disorder that causes around 100,000 deaths per year. One out of four people who have a PE will die without warning and 10 to 30 per cent of people die within the first month of diagnosis. Therefore, diagnosis and immediate treatmen t is very important. However, manual reading of CT slices by radiologists is a laborious task. They have to go through the CT slice by slice to find a PE and sometimes they can be really small, such as sub-segmental PEs. It can take a long time and lead to clinician fatigue, which makes it all the more challenging. Developing a PE detection framework using deep learning in an automatic way will have a powerful clinical impact . Oral Presentation 6 There have been some previous works on this subject, but their performance was limited as they didn’t have the same volume of data to work with that this method does. Recently, PENet has shown some very promising results, but its dataset is on a smaller scale. This work is a large-scale research study, using more than 10,000 labelled studies , meaning it can achieve results that previous research has been unable to reach. It has a data partner that collects data from multiple hospitals and anonymizes it before passing it on, and its two-stage training strategy uses a combination of attention training and patient-level studies to provide a framework that can diagnose a PE on a patient level . Automatic Diagnosis of Pulmonary Embolism Using an Attention-guided Framework: A Large-scale Study