2142 - Real-Time UNET Auto-Segmentation on Image Latency Corrected MR Cine Imaging

Purpose/Objective(s): To evaluate real-time auto-segmentation performance of the UNET model on 2D cine imaging for a moving target in MR-Linac while compensating for latencies due to image acquisition and beam delivery.

Materials/Methods: In this study, three liver (160x160 pixels) and one kidney (80x80 pixels) 2D cine image datasets with pixel spacing of 1.2mm were used, collected on Philips Ingenia MRSim using an equivalent balanced sequency used in Elekta Unity MRLinac. Each dataset comprised ground truth images and the counterpart generated by our in-house developed ConvLSTM model, which predicts frame 2400ms in future from the last input image. Manual contours on ground truth images in each dataset were delineated by experts. UNET from Medical Open Network for AI (MONAI) framework was used for auto-segmentation. The contours output by UNET on both image subgroups were compared to the manual contours in ground truth image subgroup per dataset. Dice similarity coefficient (DSC), 95th-percentile Hausdorff Distance (HD95), and center-of-mass errors (COMEs) were selected as evaluation metrics. Equivalence testing was then performed using two one-sided tests (TOST) between metrics of ground truth and ConvLSTM-generated images in each dataset. Computations were carried out using one A100 GPU and two ROME CPUs.

Results: Auto-segmentation of target was achieved within 5ms per cine frame. DSC of ground truth and ConvLSTM-generated images were (0.979±0.004, 0.966±0.020), (0.988±0.001, 0.970±0.024), (0.989±0.002, 0.961±0.038), and (0.987±0.002, 0.972±0.015) with the equivalence margin of 0.02, 0.03, 0.04, and 0.02 for Kidney, Liver 1, Liver 2, and Liver 3 Dataset, respectively. The HD95 of ground truth and ConvLSTM-generated images were (1.249±0.176 mm, 1.714±0.973 mm), (1.718±0.293 mm, 3.697±2.578 mm), (1.535±1.545 mm, 4.557±3.523 mm), and (2.370±0.807 mm, 4.243±1.728 mm) with the equivalence margin of 0.64, 2.42, 3.6, and 2.17 mm for Kidney, Liver 1, Liver 2, and Liver 3 Dataset, respectively. Regarding COMEs of ground truth and ConvLSTM-generated images, Kidney, Liver 1, Liver 2, and Liver 3 Dataset attained scores of (0.261±0.396 pixels, 0.594±0.632 pixels), (0.598±0.488 pixels, 2.174±2.102 pixels), (0.445±0.392 pixels, 2.403±2.820 pixels), and (0.578±0.482 pixels, 2.216±1.724 pixels) with equivalence margin of 0.46, 1.94, 2.39, and 1.91 pixels respectively.

Conclusion: Our study demonstrated the ability of UNET to auto-segment 2D MR cine imaging in real-time. Additionally, the equivalence test demonstrated the clinical feasability of using the UNET model with ConvLSTM image prediction model for compensating for image acquisition and beam delivery latencies.