Room: Stars at Night Ballroom 2-3
Purpose: Tissues in the head fall into 3 broad intensity ranges (air, soft tissue and bone). In order to better preserve contrast in these tissues, a Convolutional Neural Network (CNN) was developed to train synthetic CT (MRCT) images for 3 separate tissue groupings.
Methods: A U-Net architecture was developed for training. The network consisted of 14 convolutional layers. Transpose convolution was used to perform upsampling. The input data consisted of 9192 sagittal slices of T1-weighted image volumes from 47 patients. Three separate output channels were simultaneously trained, representing â€œairâ€? (<-100 HU), soft tissue (-100 to 100 HU), and â€œboneâ€? (>100 HU) regions of the corresponding CT images. MRCT images were assembled by combining images from the three channel outputs. 5-fold cross validation was applied during training.The trained network was tested on images from 10 new patients. Regions of interest (ROIs) containing 1) soft tissue and 2) bone were extracted automatically from the CT images. Mean Absolute Error (MAE) was calculated within these ROIs. Alignment of targets from MRCT and CT to Cone Beam CT (CBCT) images acquired during treatment for 6 of the test patients, were compared.
Results: MRCT volumes preserved the appearance of sulci and other soft tissue features in the brain. The average MAEs were 6.6 (range 4.1 to 13.1) and 163.1 (range 148.2 to 190.4) HU in soft tissue and bone, respectively, demonstrating reasonably good agreement with CT. Alignment differences using MRCT versus CT to CBCT images yielded standard deviations of 0.2,0.3,0.3 mm about the LR, AP, and CC axes respectively.
Conclusion: Training a CNN with three separate output channels, representing the major contrast regions of tissue in the head, yields MRCT volumes with high fidelity of contrast preservation in soft tissue and overall accuracy that is acceptable for clinical use in Radiation Therapy.
Funding Support, Disclosures, and Conflict of Interest: Supported by NIH R01EB016079