Room: Exhibit Hall | Forum 9
Purpose: Cone-beam computed tomography (CBCT) has been widely applied in medical imaging because of the low dose, high spatial resolution. Moreover, CBCT has extremely mechanical accuracy requirements on the ray source, detector and scanned object for geometry and mounting accuracy because of the use of FDK reconstruction algorithm. Therefore, geometrical correction methods is commonly used for CBCT. Recognizing the bearing ball (BB) calibrator in the projection image is the core step of geometric correction method. And we proposed a method for BB recognition based on deep learning.
Methods: The proposed method combined U-net and mask R-CNN to recognize the BB in projection image. Specifically, the method consists of six steps: 1), the original projection image is preprocessed by gauss filtering and un-sharp masking enhancement. 2), the preprocessed image is divided into several blocks. 3), the mask images of BB are segmented by U-net. 4), the image blocks including only BB are obtained by multiplication of preprocessed image and mask image. 5), Mask R-CNN is applied for recognizing central coordinates of each BB especially for overlapping region. Finally, determining all central coordinates to three two-dimensional phantoms by using ellipse fitting. Therefore, the central coordinates of each phantom i and the BB mask are obtained. To evaluate the proposed method, both qualitative and quantitative studies were performed on simulated and realistic phantom data.
Results: We test the performance of our method on simulated and realistic phantom data. In phantom experiments with real DBT machine, the results of geometric parameters correction by our method are much closer the ground truth than the results of direct linear transformation (DLT) common method
Conclusion: In this work, the proposed integrated method, which has been tested in simulation system and the realistic systems, shows great potential for calibration correction method with the bearing ball as the calibrator.