709 Views
SCMR/ISMRM Co-Provided Workshop
SCMR 22nd Annual Scientific Sessions
Hui Xue, PhD
Staff Scientist
National Heart Lung and Blood Institute (NHLBI), National Institutes of Health
Rhodri Davies, MD, PhD
Cardiology imaging fellow
Bart's heart centre
David Hansen, PhD, MSc
Lead Developer
Gradient Software
Marianna Fontana, MD
Consultant cardiologist
National Amyloidosis Centre, University College London
James Moon, MD
Clinical Director, Imaging
Barts Heart Centre and UCL
Peter Kellman, PhD
Senior Scientist
NHLBI, NIH
Background:
Deep-learning based AI algorithms show great potential to improve CMR imaging and analysis [1]. While training is achieved "off-line" with saved data, applying models, inference, will benefit from "in-line" workflow where results are immediately available after data acquisition. Thus, we extend the well-known Gadgetron framework to provide "Inline AI" functionalities for effective model inference and demonstrate it on two clinical applications: inline cine segmentation and perfusion AIF detection. Both are deployed to hospitals for clinical validation.
Methods:
Two model inference schemes are supported: a) Python Gadget supplied by user where AI models are loaded in initialization and applied to incoming data; b) Python/C++ interface where user supplies model loading and inference called in C++ runtime directly. Both schemes were implemented in Gadgetron (3.17.0, Inline AI release) and utilized in two AI applications. Cine using Python Gadget: Retro-gated cine imaging was reconstructed in Gadgetron and fed into the pre-trained CNN model [2] for myocardium segmentation. TensorFlow was used and editable endo/epi contours were sent back to scanner (Fig. 1). Perfusion using Python/C++interface: CNN model was trained to detect LV blood pool of arterial input function (AIF) image series during adenosine stress perfusion. PyTorch was used and LV AIF signal and perfusion flow maps are generated fully automatically and displayed in-line on scanner (Fig. 2). NIH OHSR (Exemption #13156) approved anonymized data processing.
Results:
Both applications were deployed to hospitals and fully "in-line" to achieve seamlessly integration with clinical MR scans. Model loading took approx. 350ms for cine and 100ms for perfusion AIF detection. Model inference for cine took ~100/25ms per image on cpu/gpu. Perfusion took ~90/10ms on cpu/gpu.
Conclusion:
Gadgetron was extended to support Inline AI inference for main-stream DP packages (TensorFlow and PyTorch) and flexible deployment schemes. Two clinical AI applications were developed and deployed, demonstrating capacities of Inline-AI Gadgetron.