A Weakly-Supervised Approach for Automatic Selection of Uniform Regions in Clinical CT Images
Presentations
SU-F-201-1 (Sunday, 7/10/2022) 2:00 PM - 3:00 PM [Eastern Time (GMT-4)]
Room 201
Purpose: Many methods for real time analysis of clinical image quality rely on identification of uniform regions of interest (UROIs) in patient images. Deep learning approaches are well-suited for automating this task, but conventional supervised learning requires time-consuming manual annotation of UROIs. This work demonstrates a weakly-supervised approach for selecting UROIs without the need for manual annotations for training.
Methods: We hypothesized that the surrogate task of distinguishing between patches of pure CT noise and random patches of patient images is sufficient for training a CNN to identify UROIs. A CNN was optimized to perform binary classification using 24 abdominal CT exams with corresponding low-dose noise simulations. The patients were randomly divided into training (n=12) and testing (n=12) sets. The training data was split into two surrogate classes: 1) 100,000 random patches from soft-tissue regions in the routine-dose exams; 2) 100,000 random noise patches from the difference between low-dose and routine-dose images. The mean and standard deviation of the patches were normalized to prevent these features from being used for classification. For validation and testing, 586 patches were randomly cropped from the patient images in the test set and then manually annotated as being either positive (n=135) or negative (n=451) examples of UROIs. The CNN performance was evaluated by calculating the area under the receiver operating characteristic (ROC) curve and inspecting heatmaps of prediction scores on the reserved test images.
Results: The area under the ROC (AUC) for the detection of uniform patches was 0.94.
Conclusion: The proposed weakly-supervised method is sufficient for optimizing a CNN to identify UROIs in patient images. Since noise patches can be readily obtained through phantom scans or simulation, this could enable faster and more robust patient-specific image quality assessments based on UROIs.
Keywords
CT, Image Analysis, Computer Vision
Taxonomy
IM- Dataset Analysis/Biomathematics: Machine learning
Contact Email