Computer Vision News - March 2019

14 Computer Vision News Challenge • We learned that the submitted algorithms generally performed quite well in assessing cancer cellularity for H&E breast cancer tumor patches with the majority of submitted algorithms having a predication probability value (p k ) greater than 0.90 on a scale of 0.0 to 1.0. This is somewhat comparable to the human pathologist readers who achieved predication probability values of 0.96 and 0.93 on our test dataset. While algorithm performance still needs to be confirmed on a much larger and more diverse patient’s dataset to verify these results, the challenge suggests that automated cancer cellularity scoring may be a reasonable approach to consider in order to reduce variability among pathologists and potentially streamlining the assessment of residual cancer burden in breast and other cancers. • We learned that a well-organized challenge in the medical field requires people of many expertise working together. For this challenge, we had experts in the clinical task (to define the task to study and its clinical relevance), statistical analysis (to determine the evaluation of the results), logistics support (to ensure that the challenge runs smoothly), as well as the support of organizations for this challenge (to advertise the challenge, and have a venue for reporting the results). Without these experts working together, the challenge would not have been successful. We have collected comments by the two winning teams. David Chambers , Senior Research Engineer at the Southwest Research Institute (SwRI) : “ The cellularity is a measure of the area of the patch that is occupied by malignant cells, so the problem can be thought of as a segmentation problem underneath. The Southwest Research Institute and Univ. of Texas Health Sciences Center at San Antonio team approached the problem as a weakly- labeled segmentation problem, and iteratively refined our algorithm by providing strong labels for hard examples and retraining. Pathologist expertise played a large role in our success. Because global context is important in this problem, we used a network with "Squeeze-and-Excitation" architectural units, a recent development in neural network architectures that allows for reweighting features according to global content. ” Co-winner Mamada Naoya , a master's course student at the Tokyo Institute of Technology , concludes: “ I study deep learning applications for material science and I know next to nothing about pathology and cancer. So my winning shows the versatility of deep learning, I believe. And I was surprised to hear that top models' performances are as good as expert pathologists'. With larger amount and more variety of data (annotation by many pathologists, different optical devices, rare clinical cases, etc.), super-human models will be possible. " Challenge David Chambers Mamada Naoya