Primary and adaptive resistance to checkpoint blockade in lung cancer

Abstract

In CT-based lung cancer screening and incidentally detected indeterminate pulmonary nodules, radiomics has shown value in improving diagnostic accuracy to discriminate cancer from benign pulmonary nodules. We have analyzed data from the National Lung Screening Trial, NLST, to identify subjects with cancerous and benign nodules, and have organized them into cohorts based on their screening history. Patients who are diagnosed with cancer following a prior nodule-negative screen have significantly worse outcome than patients who develop cancer following a prior nodule-positive/ cancer-negative screen. In cohorts of patients with nodules not diagnosed as cancer in the first screen, we have identified significant radiomic features that can discriminate those who will subsequently develop cancer from those that remain benign with an accuracy of 80%, and this can be the basis for a “radiomics risk score” to predict subsequent cancer development. However, even the seemingly large “big data” NLST data set is underpowered once cohorts are assembled with similar histories, and co-variates are accounted for. Solutions to the problem of generating sufficiently powered data sets include capturing the radiomic data at the point of care (i.e. by the radiologists) and inter-institutional sharing of images, data, features and algorithms, which have yet to be reliably implemented.