Purpose: The need for reporting patient CT dose has now focused on methods to predict patient-specific organ doses instantly and accurately. This study is an attempt to address a technical barrier - the lack of tool for automatic multi-organ segmentation in CT images - using the latest deep-learning techniques followed by realtime GPU dose computing.
Methods: Deep convolutional neural network (dCNN) for organ segmentation is trained to automatically delineate radiosensitive organs from CT. More than 30 full-body patient voxel phantoms and several hundreds of cadaver full-body CT scan datasets, containing segmented organ information are used to generate both anatomy-specific and CT protocol-specific information as training data. The learned and tested knowledge includes comprehensive anatomical variation representing voxel-wise organ definition and CT scanning protocols. This is followed by feeding a patient-specific phantom with newly segmented organs into the GPU-based Monte Carlo dose engine to derive organ doses.
Results: The dCNN has 19 layers to process features from multiple views and multiple scales for CT image segmentation. 3D CT volumes are segmented in three views, i.e. axial, sagittal and coronal, and the results are fused together to achieve the final segmentation. Model is tested in a cohort of datasets consisting of several hundreds of CT volumes with and without prior manual organ segmentation. As many as 20 organs are segmented automatically to date, showing the feasibility of obtaining satisfactory organ segmentation Dice values. The organ dose simulations using a GPU-accelerated code took less than one second to yield, for examples, 5.3xE-7 and 6.2 xE-7 MeV/g/source particle for the segmented liver for the training and testing phantom, respectively, under an abdominal CT protocol.
Conclusion: Convolutional neural network (CNN) models allow organs of interested to be segmented accurately and efficiently from patient-specific CT for dose reporting.
Funding Support, Disclosures, and Conflict of Interest: R01EB015478 and R42EB019265-01A1 (via Virtual Phantoms, Inc)