Deep learning based methods have achieved promising results for CT metal artifact reduction (MAR) by learning to map an artifact-affected image or projection data to the artifact-free image in the data-driven manner. Basically, the existing methods simply select a single window in the Hounsfield unit (HU) followed by a normalization operation to preprocess all training and testing images, based on which a neural network is trained to reduce metal artifacts. However, if the selected widow contains the whole range of HU values, the model is challenged to predict the dedicated narrow windows accurately since the contribution of small HU values to the training loss may not be sufficiently weighted relative to that for large HU values. On the other hand, if a selected window is small, the opportunity will be lost to train the network effectively on features of large HU values. In practice, various tissues and organs in CT images are inspected with different window settings. Therefore, here we propose a multiple-window learning method for CT MAR. The basic idea of multiple-window learning is that the content of large HU values may help improve features of small HU values, and vice versa. Our method can precisely process multiple specified windows through simultaneously and interactively learning to remove metal artifacts within multiple windows. Experimental results on both simulated and clinical datasets have demonstrated the effectiveness of the proposed method. Due to its simplicity, the proposed multiple-window network can be easily incorporated into other deep learning frameworks for CT MAR.
|