Room: Track 1
Purpose: To develop a widely accessible phantom-based training framework for convolutional neural network (CNN) noise reduction. Many spatially matched high and low noise (dose) CT images are typically used as training input and target respectively. Projection domain noise insertion is the current standard for generating training data by simulating low dose CT exams; however, restricted access to proprietary projection data makes it unfeasible for most institutions to implement these methods. We propose an image based framework in which phantom noise measurements are superimposed on patient data to synthesize training examples without using proprietary information.
Methods: Corrupted training inputs were generated by superimposing phantom noise measurements on patient images. Noise measurements were obtained by taking the difference between independent low dose anthropomorphic phantom scans. Medical image examples were obtained from ten routine dose patient exams. The noise and patient images were each cropped into 50,000 unique patches containing three adjacent slices (64x64x3 voxels). Prior to each training epoch, the noise patches were shuffled and added to the medical image examples to generate corrupted images to be used as the training input; this stochastic process resulted in 2.5 billion possible corrupted image realizations for training. A residual architecture was used containing 18 layers, each with 128 features, and mean-squared-error loss.
Results: When the CNN was tested on five reserved quarter dose images, standard deviation noise at the aorta was reduced by 63±4 % and normalized-mean-square-error (NMSE) was reduced by 50±5 % with respect to routine dose. Visual analysis and line profiles indicated the method significantly reduced noise while maintaining spatial resolution of anatomic features and noise texture.
Conclusion: The proposed phantom-based training framework demonstrated extensive noise reduction while maintaining spatial resolution. Because this method is based within the image domain, it can be easily implemented without access to proprietary projection data.