8 Intelligent localisers: an integrated time-saving deep learning solution for the planning of cine imaging and identification of unexpected findings from a single transaxial stack. (1st November 2021)

• Record Type: Journal Article
• Authors: Howard, James; Zaman, Sameer; Francis, Darrel; Cole, Graham
• Is Part Of: Heart. Volume 107(2021)Supplement 3
• Page Start: A5
• Page End: A6
• Publication Date: 2021-11-01
• Subjects: Heart -- Diseases -- Treatment -- Periodicals; Cardiology -- Periodicals; 616.12
• DOI: 10.1136/heartjnl-2021-BSCMR.8
• Languages: English
• ISSNs: 1355-6037

Abstract

Abstract : Background: CMR scans typically begin with a transaxial stack for anatomical evaluation, but current strategies require (1) additional localiser images for cine planning and (2) expert staff present to identify unexpected findings and adjust protocols for maximum yield. We present an integrated deep learning solution to (1) automate scan plane prescription without need for localisers and (2) identify important protocol-changing findings at the earliest stages of the scan. Methods: Task 1: We trained a neural network on 1, 300 CMR transaxial stacks to identify the spatial co-ordinates of the cardiac apex and the 6 locations on the mitral annulus corresponding to 2-, 3-, and 4-chamber views. By identifying these points, the network proposed long-axis scanning planes, which were compared with a gold-standard of experienced radiographers who had access to the full conventional range of localiser images. 130 scans were used as a testing set. Task 2: We trained a neural network on 1, 500 transaxial stack slices to perform automated segmentation of important cardiac structures, great vessels, and other pathologies. The system assembled a 3D model of the thorax from which it made clinical measurements which were compared to those obtained from the full scans. 200 scans were used as a testing set. Results: Task 1: The network successfully learned the 7 key points on the transaxial stack, and the scan planes it proposed closely matched those chosen by radiographers using dedicated localisers, with a median out-of-plane distance at the key points of only 3.4 mm (IQR 1.6–6.3) - less than half the slice thickness (8 mm) of the gold standard cines. Task 2: The system was successful in segmenting the transaxial slices (Dice 0.91). Diagnostic accuracy was high for abnormalities including left and right ventricular dilatation (90.5%, 85.5%), LVH (85%) and ascending aorta dilatation (94.4%). The ROC AUC for diagnosing pleural effusions was 0.91. Conclusion: A neural network can use the standard transaxial stack of images to identify important pathology and accurately predict scan planes close to those of expert radiographers using dedicated localisers. This system could speed up the delivery of CMR and lead to faster and more dynamic scanning protocols.