(Wednesday, 7/17/2019) 10:30 AM - 11:00 AM
Room: Exhibit Hall | Forum 2
Purpose: To develop a methodology for automatically identifying the anatomical location of CBCT scans from single projection images using an AlexNet based deep-learning image classifier.
Methods: Anonymized clinical CBCT projection data were collected from 4914 scans over 14-months from two LINAC-mounted CBCT systems. Scans were manually labeled as ‘Head’, ‘Neck’, ‘Thorax’, ‘Abdomen’ or ‘Pelvis’ based on the mid-plane of the scanned volume relative to standard anatomic features. Anteroposterior (AP/PA) projection images were used after downsampling to 256x256. The training and testing datasets used for the classifier were separated by 6 months to ensure patients on treatment had completed their treatment resulting in 3288 projections used for training and validation and 2585 projections for testing. Chi-squared test and Cramer’s V were used to determine significance and strength of association between the categorization methods.
Results: The categorization of the time-separated testing data resulted in per-class accuracy of 95.3%, 90.5%, 92.8%, 72.4% and 94.4% for ‘Head’, ‘Neck’, ‘Thorax’, ‘Abdomen’ and ‘Pelvis’ respectively. The majority of misclassifications were assigned to a neighboring anatomical site (i.e., “near misses�) particularly when the scan mid-plane was near the dividing line between categories. The low performance in the ‘Abdomen’ may have been related to the lack of distinct anatomic landmarks and/or low x-ray contrast in this site. The chi-squared test between automatic and manual categorization showed significance (p<0.0001). Cramer’s V, showed the strength of association was strong with a value of 0.854. The computation times for training the network and categorizing the 2584 image test dataset were 270s and 70s respectively.
Conclusion: This study indicates that anatomical sites can be accurately determined with deep-learning neural networks using single projection images from AP/PA views. The 27ms per image average categorization rate opens the possibility of use in real-time decisions, such as aiding in the automation of scan protocol selection.
Funding Support, Disclosures, and Conflict of Interest: Funding was provided in part by NIH T32EB002103 and Varian Medical Systems. Hania Al-Hallaq receives royalties and licensing fees for computer-aided diagnosis technology through the University of Chicago.
Keywords
Image-guided Therapy, Cone-beam CT
Taxonomy
IM/TH- RT X-ray Imaging: CBCT imaging/therapy implementation
Contact Email
