Stability of artificial intelligence methods in computed tomography metal artifacts reduction task

Reducing image artifacts in modern computed tomography is one of the most important directions, which also receives a research boost from the rapid development of artificial intelligence and machine learning. The neural network based methods often allow the achievement of the high artifact reduction quality unattainable by classical approaches. Simultaneously, these models tend to be less robust and generalize worse than algorithmic approaches. The output result may degrade severely under the use in cases not represented in learning datasets. In the present work, the robustness and generalization properties of neural network models are studied in the task of metal-like artifact reduction for computed tomography reconstruction. The pipeline for data augmentation is developed for the metal artifacts severity and geometry variations parameters. The pipeline is validated by augmentation of DICDNet model test dataset and evaluating of the corresponding pre-trained DICDNet model. The generated data consisted of 192000 samples for 12×8 combinations of experiment parameters. It was demonstrated that DICDNet is reasonably robust to geometry variations with a maximum deviation of PSNR≈1.5%. At the same time, even small artifact severity growth leads to rapid quality degradation and blurring of studied model output. The degradation reaches ≈22% PSNR deterioration from the baseline algorithmic approach, which is also one of the inputs for the studied model.