SHU-Aug: Stochastic Hounsfield Unit Augmentation for Clinical Target Volume Segmentation of Glioma

The accurate segmentation of clinical target volumes (CTVs) plays a vital role in radiotherapy planning, but is time consuming, labor intensive, and prone to inter-observer variation. Automated CTV segmentation contributes to both the efficiency and quality of segmentation. Recently, deep neural network (DNN)-based segmentation methods have achieved remarkable improvements. However, a CTV can contain tumor spread tissues and subclinical disease tissues, resulting in a poorly defined border and irregular shape. We hypothesize that leveraging the imaging characteristics of computed tomography (CT) in DNNs will improve the accuracy of CTV segmentation. In CT images, the Hounsfield unit (HU) scalar represents the absorption rate of X-rays by tissues, providing valuable information for CTV segmentation. This paper proposes the stochastic HU augmentation (SHU-Aug) method, which aims to enhance the variability in CT data. Instead of using constant minimum and maximum HU values during the normalization of the CT image, SHU-Aug converts the value from a constant to a variable that obeys a distribution, e.g., a uniform distribution or normal distribution. The proposed SHU-Aug for CT is analogous to color augmentation in natural images, which has been frequently used in the training of DNNs. Extensive experiments on various DNN architectures, including U-Net, Attention U-Net, and DeepLabV3+, demonstrate the advantages of the proposed SHU-Aug method.