Practical applications of deep learning: intelligent image hanging protocol for abdominal magnetic resonance imaging

Background It is time-consuming to open an abdominal MR in traditional PACS and put all the image serials in the proper order before the radiological diagnosis. In this study, we aim to develop and validate an intelligent tool to assist radiologists in hanging abdominal MR images before radiological diagnosis. Methods Two independent cohorts were utilized in this study. The developing cohort included abdominal MR images of 1374 patients randomly collected from four centres, while the clinical evaluation cohort included images of 481 consecutive patients from one centre. A series of deep learning algorithms and rules were built to implement image preprocessing, sequence classification, and optimum sequence selection, which together enable full process automation for hanging images. The system was evaluated from two aspects: i) accuracy of discriminating MR sequences and phases and ii) performance in real clinical scenarios, including accuracy, applicability, and efficiency. Results The model had high accuracy in discriminating 13 common types of MR sequences and phases (the average accuracy was 99.1% and ranged from 96.2–100%). In clinical evaluation, the model successfully covered 98.5% of patients, and 85.9% of these achieved 100% accuracy in image alignment. Multivariate logistic regression analysis revealed that none of three factors, including contrast agent type, MR device, and liver background showed statistical significance as factors associated with model mistakes. With the assistance of this tool, the time spent on hanging images dropped from an average of 118.2 s per case to 22.2 s. Conclusion This intelligent tool can be used to assist radiologists in hanging abdominal MR images, reducing their workload, and improving efficiency.