Monday, December 19, 2016
Robust to the future of medical imaging
Machine learning can make medical image feature detection, annotation, segmentation, and classification much more robust to changes in the lab and clinic. Traditional deterministic image processing uses fixed pipelines of filters to denoise, simplify, and enhance. The deterministic filter pipelines are designed by experts in image processing who have spent time looking closely at the images, understanding the imaging system that took the images and understanding the goals of the imaging. Also any deterministic feature extraction process will always look at finite number of image features - even if that number is in the hundreds. What happens when looking at a new condition or disease that affects new never before examined features? What happens when the imaging process changed? When a dye is changed? When a MRI pulse sequence is changed? The deterministic pipeline fails in these cases and the pipeline itself has to be changed - meaning rewriting the program to some degree. Machine learning image processing systems only need retraining instead of rewriting making them more robust to future imaging and medical advancements and new areas of research.