HDR UK Applied Analytics Webinar: Deep Learning Methods on Medical Imaging - Grant Mair, Wenwen Li, Alessandro Fontanella (55:17)

The Challenges and Applications of Deep Learning Methods on Medical Imaging for Acute Ischemic Stroke - With huge amounts of brain CT data generated daily in routine clinical practice, there is great potential to harness the predictive power of machine learning (ML) and especially deep learning (DL). However, developing DL methods for stroke, especially acute ischemic stroke (AIS) is challenging for several reasons. First, access to imaging data for research is rightly restricted, rarely centrally stored, or standardised. Second, gold-standard expert human interpretation of the imaging is qualitatively assessed by radiologists, highly variable and recorded in free text, while a quantitative ground-truth is rarely achievable.

Third, standardising clinical brain CT data for DL in terms of quality and format is a highly time-intensive process with very limited prior research specific for AIS, and no fully open-source code available. Fourth, the black-box nature of DL makes the identification and evaluation of reliable biomarkers more problematic than expected. In this talk, we will show the gap between routine clinical CT image quality and the data quality requirements for ML or DL. We will describe our bespoke pipeline for processing routinely-acquired clinical CT for DL methods from a large international trial dataset – The Third International Stroke Trial (IST-3). Then we will present our novel DL method for learning AIS lesion patterns on CT and discuss its efficacy and clinically relevant observations. Finally, we will share examples of our attempts to visualise the regions of interest detected by our DL method as it attempts to identify stroke lesions and explore the potential and challenges of making meaningful interpretations from DL. (Grant Mair, Wenwen Li, Alessandro Fontanella) 19/01/2022