Evaluation of an Automatic Classification Algorithm Using Convolutional Neural Networks in Oncological Positron Emission Tomography

Abstract

Introduction: Our aim was to evaluate the performance in clinical research and in clinical routine of a research prototype, called positron emission tomography (PET) Assisted Reporting System (PARS) (Siemens Healthineers) and based on a convolutional neural network (CNN), which is designed to detect suspected cancer sites in fluorine-18 fluorodeoxyglucose (18F-FDG) PET/computed tomography (CT).Method: We retrospectively studied two cohorts of patients. The first cohort consisted of research-based patients who underwent PET scans as part of the initial workup for diffuse large B-cell lymphoma (DLBCL). The second cohort consisted of patients who underwent PET scans as part of the evaluation of miscellaneous cancers in clinical routine. In both cohorts, we assessed the correlation between manually and automatically segmented total metabolic tumor volumes (TMTVs), and the overlap between both segmentations (Dice score). For the research cohort, we also compared the prognostic value for progression-free survival (PFS) and overall survival (OS) of manually and automatically obtained TMTVs.Results: For the first cohort (research cohort), data from 119 patients were retrospectively analyzed. The median Dice score between automatic and manual segmentations was 0.65. The intraclass correlation coefficient between automatically and manually obtained TMTVs was 0.68. Both TMTV results were predictive of PFS (hazard ratio: 2.1 and 3.3 for automatically based and manually based TMTVs, respectively) and OS (hazard ratio: 2.4 and 3.1 for automatically based and manually based TMTVs, respectively). For the second cohort (routine cohort), data from 430 patients were retrospectively analyzed. The median Dice score between automatic and manual segmentations was 0.48. The intraclass correlation coefficient between automatically and manually obtained TMTVs was 0.61.Conclusion: The TMTVs determined for the research cohort remain predictive of total and PFS for DLBCL. However, the segmentations and TMTVs determined automatically by the algorithm need to be verified and, sometimes, corrected to be similar to the manual segmentation.