Room: Stars at Night Ballroom 2-3
Purpose: Develop a Deep Convolutional Neural Network for robust and fast automatic delineation of cardiac substructures in non-contrast CT images and evaluate the performance on cross-institutional data in the Radiotherapy Comparative Effectiveness randomized trial.
Methods: The model consists of two concatenated 3D U-Nets: the localization U-Net that performs context aggregating and the segmentation U-Net that performs multi-class classification. In the localization U-Net, a dilated 5Ã—5Ã—5 convolution is used to capture location information from large receptive fields. The initial number of 48 filters is doubled with pooling operations, resulting in 384 feature maps at the bottom of the U-Net. Bounding boxes of substructures are produced by the softmax classification layer. In contrast, the input of the segmentation U-Net is a two-channel volume involving the outputs of localization U-Net and the original CT images. The convolutional window size is set to 2, and the initial number of filters is halved, resulting in 192 feature maps. The training and cross-validation data include 130 sets of non-contrast CT images and the independent test involves 30 cases. To evaluate the model performance, our auto-segmented contours were compared with the expert contours using Dice similarity coefficient and the average Hausdorff distance. The expert contours were manually performed by an experienced sonographer with physician over-read using validated cardiac atlases.
Results: Between expert's and auto-segmented contours, Dice of 0.956Â±0.015, 0.904Â±0.022, 0.834Â±0.047, 0.819Â±0.049, 0.822Â±0.044 and 0.315Â±0.073 for the whole-heart, LV, RV, LA, RA and LAD were achieved in 30 independent test cases. The average Hausdorff distances were 1.5Â±0.7mm, 1.9Â±0.5mm, 2.5Â±0.7mm, 2.5Â±0.7mm, 2.3Â±0.6mm and 6.4Â±3.2mm respectively. For a new dataset, the average time to generate the contours was 45 seconds.
Conclusion: A multi-stage 3D U-Net was developed and evaluated for cardiac substructure auto-segmentation in non-contrast CT images, demonstrating state-of-the-art accuracy and fast model deployment in new cases within one minute.
Funding Support, Disclosures, and Conflict of Interest: Dr. Lei Dong is in the Speaker Bureau for Varian Medical Systems. Dr. Taoran Li has a speaker agreement with Varian Medical Systems unrelated to this work.
Not Applicable / None Entered.