| Popis: |
The MR-Linac can enable real-time radiotherapy adaptation. However, real-time image acquisition is restricted to 2D to obtain sufficient spatial resolution, hindering accurate 3D segmentation. By reducing spatial resolution fast 3D imaging is feasible. Our study evaluates how much the spatial resolution of MR-images can be reduced without compromising a deep-learning segmentation performance. We also assess the effect of an auxiliary task of simultaneous restoration of high-resolution images. Artificially downsampled images were created from 163 3D MR-scans using a k-space truncation to spatial resolution levels of approximately 1.5x1.5x2/1.5x3x3/1.5x6x6/2.5x9x9 mm3. Data was split into train/validation/test of 116/12/35 images. A U-Net was trained to obtain high-resolution segmentation of prostate, bladder, and rectum on each resolution level. Images acquired with low resolution were obtained by scanning 10 male healthy volunteers, 4 series per subject with a gradually decreasing spatial resolution. The networks trained from the artificially downsampled data were fine-tuned to segment these images. For the artificially downsampled images results in terms of DICE when including the auxiliary task, going from training on the highest resolution images to the lowest, where 0.87/0.86/0.86/0.85, 0.95/0.94/0.93/0.93 and 0.82/0.84/0.83/0.80 for CTV, bladder, and rectum, respectively while for the acquired low-resolution images 0.78/0.76/0.71/0.59, 0.82/0.85/0.81/0.74 and 0.81/0.82/0.69/0.58. The segmentation results when training including the auxiliary task were comparable for the artificial images and with a trend towards better for the acquired LR images. These results indicate that some speed-up in image acquisition can be obtained without significantly reducing the accuracy of a segmentation deep-learning network. |
