Technical Artifact Detection in CBCT Imaging by Binary Image Classification
Presentations
PO-GePV-M-28 (Sunday, 7/25/2021) [Eastern Time (GMT-4)]
Purpose: To train a deep learning TensorFlow model on cone-beam computed tomography (CBCT) images with various technical artifacts using transfer learning and to label image quality afterward as acceptable (flawless) or non-acceptable (somehow destructed).
Methods: A pre-trained 101-layer variant of the Residual Network (ResNet v2) was used to generate a deep neural network (DNN) model with custom CBCT images to classify image quality one of two categories. The first category was employed by 2081 image slices from various body regions without artifacts (except small distortions created by marker seeds). The second category included 400 images with three types of artifacts observed during clinical CBCT imaging: organ motion, ring artifacts, and unknown types. Reference images were not further sub-divided, e.g., by tube parameter or region of interest (ROI).
Results: Training for the configuration above took about 45 minutes (on an i7-10750H). The ratio between training and verification dataset was 80:20, and the evaluated internal accuracy was better than 99%. Prediction of external images took less than 3s, and all technical artifacts are assigned, while undistorted images and those with marker seeds are labeled as flawless. The algorithm typically predicts a label score of 97% or higher for those images.
Conclusion: Transfer learning was successfully applied to detect technical artifacts in CBCT imaging for an O-ring gantry-type linear accelerator. For other imaging modalities like CTs in treatment planning, CNN and algorithms can be easily adopted. Future workflows should include automated image processing to reduce potential errors for patient position matching between treatment plan geometry and daily cone-beam CT.
Keywords
Not Applicable / None Entered.
Taxonomy
Not Applicable / None Entered.
Contact Email