Purpose: The recently-introduced MR-Linac enables MRI-guided online adaptive radiation therapy (MRgOART), for which fast and accurate segmentation of the gross tumor volume (GTV) is essential. This work aims to develop a framework allowing fully-automatic segmentation of GTV based on multi-parametric MRI (mpMRI) using deep neural networks.
Methods: The framework includes (1) registering mpMRIs and delineating target organs (e.g., pancreas), (2) pre-processing mpMRIs (e.g., bias correction, normalization, resampling), (3) extracting 3D patches in the target organ as inputs for two convolutional neural networks, (4) determining the probability of each voxel belonging to the tumor, and (5) post-processing the probability map to generate the GTV. The use of this framework was demonstrated with multi-phase DCE T1 MRIs (pre-arterial, arterial and venous) acquired from 33 pancreatic cancer patients, with datasets from 28 patients for model training and 5 patients for model testing. The model-generated GTV were compared with the manual contours by experienced radiologist and radiation oncologists based on dice similarity coefficient (DSC), Hausdorff distance (HD), and mean distance to agreement (MDA).
Results: The mean values and standard deviations of the performance metrics of the 5-fold cross-validation on the training dataset were: DSC = 0.66Â±0.11; HD = 10.1Â±6.6 mm; and MDA = 2.7Â±1.7 mm. The performance of the trained model on the test dataset were: DSC = 0.71Â±0.09; HD = 7.3Â±2.8 mm; and MDA = 2.2Â±0.6 mm. The time required to generate a GTV contour was 64Â±19 sec using a Core i7 CPU and a GTX 1060 GPU.
Conclusion: A framework for auto-segmentation of GTV using deep learning algorithms based on multi-parameter MRIs was developed and was tested for pancreatic GTV based on multi-phase DCE-MRI. Efforts are underway to make the framework robust for MRgOART by incorporating larger datasets.