Subspecialty-Level Differential Diagnoses by Machine on Clinical Brain MRI

Speaker: James Gee, PhD

Date: February 28

Time: 4:00 pm

Location: SMPH/HSLC, Room 1345

Abstract: I will describe work that is the first to demonstrate expert-level diagnostic performance of an automated system for complex multi-sequence brain MRIs.  For the 19 diseases of the cerebral hemispheres under consideration, diagnostic performance was comparable to academic neuroradiology attending physicians with extensive subspecialty experience, and significantly better than general radiologists or trainees.  Importantly, the system simultaneously provides quantitative, un-biased assessments of lesion location, signal, and other important properties, which are essential pieces of treatment-relevant information to healthcare practitioners, beyond the differential diagnoses.  Moreover, because it inherently incorporates radiological knowledge, the system is capable of diagnosing both common and rare diseases, without requiring training data on each diagnosis.  Finally, our system explicitly creates a probability-ranked differential diagnosis, thereby standardizing the language of radiology reporting into the language of probabilistic reasoning.  The understanding and methodologies derived from this work could naturally be extended to any other organ system and modality within radiology.  We manage to accomplish all of this through the use of widely available classical techniques in image analysis and machine learning, a development that challenges – and I hope will stimulate discussion about – the current direction in the field.