Deep Learning Based Cine-MR Image Prediction and Analysis for Abdominal Motion
J Weng1*, S Bhupathiraju2, T Samant3, A Dresner4, J Wu1, S Samant1, (1) Department of Radiation Oncology,University of Florida, Gainesville, FL (2) Department of Computer & Information Science & Engineering, University of Florida, Gainesville, FL (3) UF Research Computing, University of Florida, Gainesville, FL (4) Philips Healthcare
Presentations
WE-C930-IePD-F5-3 (Wednesday, 7/13/2022) 9:30 AM - 10:00 AM [Eastern Time (GMT-4)]
Exhibit Hall | Forum 5
Purpose: IGRT in abdomen can be challenging in the presence of organ motion, ranging from the mostly respiratory motion of dome of liver to more complex peristalsis and respiratory motions of stomach region. MR-linacs offer real-time internal anatomy based gating but there is latency between acquired images of patient states before and during beam delivery. Here we compensate for this latency by presenting a deep learning model for organ motion prediction.
Methods: Multiplanar cine-MRI was acquired every 600ms on 10 volunteers on Philips Ingenia MRsim, which is equivalent to the MR module on Elekta Unity MR-linac. A T2/T1 balanced turbo field echo (bTFE) sequence was used. A deep learning model, using 6 input images to predict future image(s), was generated with convLSTM neural network using SSIM (Structural Similarity Index) and MSE as loss functions. SSIM, MSE, NMI(normalized mutual information) and DTA(distance to agreement) were used to evaluate performance.
Results: Planar cine-MRI prediction was successful for time frames of 0.6s-3.0s. The performance of the model remained stable over the 2min of test data with variations in SSIM and MSE within 2std. Due to noise in the cine-MRI from fast acquisition, DTA here is defined that values within 10% are considered the same. For stomach, using a patient specific model, DTA of 2.4mm over 100% ROI volume and 2.0mm for 99% ROI volume were achieved. For liver, using nonpatient specific global model, DTA of 2.7mm over 100% ROI volume and 2.3mm for 99% ROI volume were achieved.
Conclusion: Deep learning models were constructed to successfully predict motions from 0.6s-3.0s for peristalsis of the stomach (patient specific model) and for respiratory motion of liver (nonpatient specific global model). In both cases, cine images predicted for 1.8s had negligible difference from the ground truth, and DTA was <2.5mm for both stomach and liver motions.
Keywords
Taxonomy
IM/TH- MRI in Radiation Therapy: MRI/Linear accelerator combined- IGRT and tracking
Contact Email