PURPOSE The recently introduced MR-Linac enables MRI-guided Online Adaptive Radiation Therapy (MRgOART) of pancreatic cancer, for which fast and accurate segmentation of the gross tumor volume (GTV) is essential. This work aims to develop an algorithm allowing automatic segmentation of the pancreatic GTV based on multi-parametric MRI using deep neural networks. METHODS We employed a square-window based convolutional neural network (CNN) architecture with three convolutional layer blocks. The model was trained using about 245,000 normal and 230,000 tumor patches extracted from 37 DCE MRI sets acquired in 27 patients with data augmentation. These images were bias corrected, intensity standardized, and resampled to a fixed voxel size of 1 × 1 × 3 mm3. The trained model was tested on 19 DCE MRI sets from another 13 patients, and the model-generated GTVs were compared with the manually segmented GTVs by experienced radiologist and radiation oncologists based on Dice Similarity Coefficient (DSC), Hausdorff Distance (HD), and Mean Surface Distance (MSD). RESULTS The mean values and standard deviations of the performance metrics on the test set were DSC = 0.73 ± 0.09, HD = 8.11 ± 4.09 mm, and MSD = 1.82 ± 0.84 mm. The interobserver variations were estimated to be DSC = 0.71 ± 0.08, HD = 7.36 ± 2.72 mm, and MSD = 1.78 ± 0.66 mm, which had no significant difference with model performance at p values of 0.6, 0.52, and 0.88, respectively. CONCLUSION We developed a CNN-based model for auto-segmentation of pancreatic GTV in multi-parametric MRI. Model performance was comparable to expert radiation oncologists. This model provides a framework to incorporate multimodality images and daily MRI for GTV auto-segmentation in MRgOART.

中文翻译：

---

使用深度卷积神经网络在多参数 MRI 中自动分割胰腺肿瘤

目的最近推出的 MR-Linac 使 MRI 引导的胰腺癌在线自适应放射治疗 (MRgOART) 成为可能，因此快速准确地分割总肿瘤体积 (GTV) 至关重要。这项工作旨在开发一种算法，允许使用深度神经网络基于多参数 MRI 自动分割胰腺 GTV。方法我们采用了基于方窗的卷积神经网络 (CNN) 架构和三个卷积层块。该模型使用从 27 名患者的 37 个 DCE MRI 集中提取的约 245,000 个正常和 230,000 个肿瘤块进行训练，并进行了数据增强。这些图像经过偏差校正、强度标准化，并重新采样为 1 × 1 × 3 mm3 的固定体素大小。训练后的模型在另外 13 名患者的 19 个 DCE MRI 集上进行了测试，模型生成的 GTV 与经验丰富的放射科医生和放射肿瘤学家基于 Dice 相似系数 (DSC)、Hausdorff 距离 (HD) 和平均表面距离 (MSD) 手动分割的 GTV 进行了比较。结果 测试集上性能指标的平均值和标准偏差为 DSC = 0.73 ± 0.09、HD = 8.11 ± 4.09 mm 和 MSD = 1.82 ± 0.84 mm。观察者间差异估计为 DSC = 0.71 ± 0.08、HD = 7.36 ± 2.72 mm 和 MSD = 1.78 ± 0.66 mm，分别在 p 值为 0.6、0.52 和 0.88 时与模型性能无显着差异。结论 我们开发了一种基于 CNN 的模型，用于在多参数 MRI 中自动分割胰腺 GTV。模型性能可与放射肿瘤专家相媲美。