Computer Vision News - May 2018

A number of ISBI 2018 participants took part in a challenge to predict the malignancy of lung nodules based on sequential CT scans. There are two categories of computer-assisted systems in radiology: one is computer- assisted detection , the other is computer-assisted diagnosis . This challenge was about the latter. Given a pulmonary nodule in a low-dose CT scan, what is the probability that this nodule is cancerous? One of the highlights of the challenge was that the top six teams all used deep learning methods and convolutional neural networks to solve the problem. Alireza Mehrtash was just pipped to the post and landed second position on the leader board. Alireza is a research scientist at Brigham and Women’s Hospital, Harvard Medical School, Boston ; and at the same time is a PhD student at the University of British Columbia, Vancouver . He speaks to us about his team’s submission. The team’s model is based on three different methods. The first two are deep learning models and the third is just a shallow model which is a logistic regression. They trained on three different data sets – two large publicly available data sets and one just data from the challenge. The first public data set is The Lung Image Database Consortium (LIDC), which is from the National Cancer Institute and is hosted on The Cancer Imaging Archive (TCIA). The second is from Kaggle , the Data Science Bowl challenge which was held last year. Alireza explains more about the training methods they used: “ We trained a CNN that we can call a virtual radiologist, which predicts the features that radiologists usually annotate under images. We trained another CNN on data from pathology. By leveraging these two kinds of CNNs we were able to get very good results. In combination with the growth of the nodules, all these three components made a very successful model. ” Unsurprisingly for machine learning and deep learning , Alireza says the hardest and most time-consuming part of the challenge was cleaning the data, organising the data and getting to a point that it was ready to train. They overcame this by using the good open- source software that is out there based on the contributions of many data scientists over the years. Alireza concludes by telling us what he takes away from the challenge: “ I think the main and the most interesting thing that I’ve learned was that we can learn features and representations by just training models on radiologists’ annotations, and then use them in conjunction with pathology to train models that are more accurate. Human and machine together can be better than either of them alone .” Alireza Mehrtash 28 Thursday ISBI DAILY Challenge BEST OF