Mannudeep Kalra

MD

Massachusetts General Hospital

Staff Radiologist, MGH

PLCO (Learn more about this study)

PLCO-366

May 29, 2018

USE OF PLCO IMAGING DATASET FOR MACHINE LEARNING PROJECTS

Our group is involved in validation and testing of several Machine learning (ML) based algorithms for imaging applications. In the proposed project, we would like to use the PLCO imaging datasets (only imaging including radiographs, CT and MR) and their labels for validation and testing of various ML algorithm at our disposal. For the project, we will randomly select imaging exams and establish gold standard using the PLCO information on findings as well as re-interpretation of findings by a set of radiologists in consensual reading. Additionally, the imaging datasets will be processed with ML algorithms to determine the ML-findings. Separate radiologists will then serve as blinded test readers. Findings from ML and test radiologists will be benchmarked against the gold standard. The ML algorithms typically provide heat maps and prediction statistics for detected radiologic findings. Radiologists will be asked to grade likelihood of radiologic findings on a five-point categorical scale (1= definitely absent, 2= probably absent, 3= uncertain, 4= probably present, 5= definitely present). Data will be analyzed using receiver operating characteristics (ROC) analyses. We may use part of the PLCO data for training and validation aspects of ML algorithms should the test suggest limitations in the tested algorithm. Some studies will involve mixing of PLCO imaging datasets with deidentified datasets from our institution to increase the heterogeneity of assessed examinations, and to add more contemporary datasets acquired on newer machines and image processing programs. We will not share the PLCO datasets with collaborators not listed in our proposal without first applying and obtaining permission for their inclusion on the NIH data access center.

Presently, we have access to several ML algorithms that assess chest radiographs and chest CT for a host of findings. The ML for chest radiograph can identify findings on chest radiographs including blunted costophrenic angle, pleural effusion, pulmonary opacity, consolidation, cavitation, emphysema, interstitial fibrosis, parenchymal calcification, cardiomegaly, hilar prominence, pneumothorax, and degenerative spinal disease. The chest CT ML algorithms can also detect several findings such as pulmonary nodules, opacities, lung fissures, emphysema, pulmonary embolism, pneumothorax, coronary artery calcification, aortic dimensions, and pleural effusions, bone density and presence of fatty liver. We anticipate working with other advanced ML algorithms including CT and MR for head, abdomen and pelvis as well. We will present and publish our work in various peer reviewed platform.

Although some have proposed that ML algorithm can replace human radiologists, we believe that in the short term, these ML algorithm will help make radiologists accurate and efficient for interpreting radiologic examinations. This can help cut costs associated with imaging.