High-throughput image labeling and quality control for clinical trials using machine learning

Background: Manually importing and analyzing image data can be time-consuming, prone to human error, and costly for large clinical trial datasets. This can lead to delays in quality control (QC) feedback to imaging sites and in obtaining data analysis results. Herein we describe the creation and application of a high-throughput review process for import, classification, labeling and QC of large multimodal clinical trial image datasets. Methods: Automated methods were used to remove patient identifying information, extract image header data, and filter image data for usability. A convolutional neural net was applied to estimate anatomy for CT images. Internal scores were assigned for each image series to identify the optimal series for labeling and reading of each anatomical region. Image QC reports were automatically generated for all patients. Results: In combined studies for which 204,492 series were received, 27,841 series were identified as usable and 13,415 series were labeled. Using this high-throughput method, total work-hours required per time point were reduced by an approximate factor of ten when compared to traditional review and labeling methods. Our anatomic classification system identified 95.7% of image series correctly, with the remaining series being manually corrected before labeling and analysis. Conclusions: A high-throughput image analysis pipeline was implemented in a large combined dataset of clinical trial image series. This pipeline can be applied across other studies and modalities for fast image data characterization, labeling and QC.