In-house development, implementation and evaluation of machine learning software for automated clinical scan processing. Issue 10 (17th May 2021)

• Record Type: Journal Article
• Authors: Taylor, Jonathan C.; Sharkey, Michael J.; Metherall, Peter
• Is Part Of: Nuclear medicine communications. Volume 42:Issue 10(2021)
• Page Start: 1157
• Page End: 1161
• Publication Date: 2021-05-17
• Subjects: artificial intelligence -- machine learning -- myocardial perfusion imaging -- single-photon emission computed tomography; Nuclear medicine -- Periodicals; 616.07575
• DOI: 10.1097/MNM.0000000000001436
• Languages: English
• ISSNs: 0143-3636

Abstract

Abstract : Objectives: Scanning of myocardial perfusion patients on a dedicated cardiac gamma camera (GE Discovery NM 530c) requires careful positioning between stress and rest acquisitions. The offset between scans is routinely measured through image registration and analysis of the transformation matrix. Accurate registration requires a 3D mask to be drawn manually over the left ventricle, excluding any significant extracardiac tracer uptake. This work sought to automate mask drawing as part of a new, more efficient system for checking relative patient position. Objectives were to (1) develop and test a machine learning algorithm for automated segmentation of stress scans; (2) implement the algorithm as a clinical position-check software package and prospectively compare performance with manual methods. Methods: Algorithm development utilised 9604 manually drawn segmentation masks (10% for validation, 10% for testing). The NiftyNet platform was used to train, optimise and test a convolutional neural network. The algorithm was packaged as a clinical tool and utilised prospectively alongside the manual technique. The software was evaluated for 343 patients to ensure adequate functioning and to assess performance. Results: The difference in patient offset measurements between manual and automated methods was small (mean of −0.01 mm (±0.4 mm) in the test dataset, mean difference of −0.05 mm (±0.5 mm) during clinical evaluation). The position-check software was found to be reliable during prospective evaluation, producing segmentations that adequately enclosed the left ventricle in all cases. Conclusion: This work demonstrates that established machine learning technology and modest hardware can be used to create automated segmentation tools that perform well in the clinic.