Room: AAPM ePoster Library
Purpose: Convolutional neural networks (CNNs) have achieved excellent results in medical image segmentation. In this study, we proposed a novel multi-path 3D Dense-UNet for generating accurate glioblastoma (GBM) GTV contours from multi-modal MR images. We hypothesized that this multi-path model could achieve more accurate segmentation than a single-path model.
Methods: 258 GBM patients were included in this study. Each patient had four multi-modal (T1-weighted, contrast-enhanced T1-weighted, T2-weighted, and FLAIR) MR images. GTV contours were delineated by experienced neuro-radiologists. We built a multi-path 3D Dense-UNet that could be trained to directly convert four MR images to a GBM GTV contour. Both single-path and multi-path models use Encoder and Decoder architectures. All four images were concatenated and fed into the same Encoder path in the single-path model. In the multi-path model, each image had its own Encoder path where features maps were fused by Squeeze-and Excitation Blocks. The patient cohort was randomly split into a training set of 180, a validation set of 39, and a testing set of 39 patients. The multi-path 3D Dense-UNet was compared with the single-path model using the Dice coefficient between the manually drawn and autosegmented contours. A Wilcoxon signed-rank test was conducted to examine the model difference.
Results: The single-path Dense-UNet and multi-path Dense-UNet achieved Dice coefficients of 0.911±0.058 (mean±SD) and 0.922±0.041, respectively. The p-value of the Wilcoxon signed-rank test was less than 0.01.
Conclusion: Our proposed multi-path 3D Dense-UNet generated more accurate GBM GTV contours than the single-path model. The statistical test indicated a significant difference between the single-path and multi-path models. The proposed multi-path technique could be integrated into image-transfer CNNs used for other applications in the future, such as those used for OAR segmentation or synthetic CT generation.
Funding Support, Disclosures, and Conflict of Interest: This study was funded by Varian Medical Systems, Inc.