Purpose: To incorporate the concept of wavelet transform into the U-net for the organ segmentation. The segmentation of organs in CT images is a crucial part of the overall treatment planning process in radiation oncology. While many automation packages, both atlas- and deep learning-based, are becoming available recently, current clinical practice for both organ-at-risks (OARs) and targets still relies almost entirely on human inputs. Here we incorporated the wavelet’s multi-level spatially encoding characteristics into the U-net, and examined the clinical performance of the resultant automated segmentation system.
Methods: A total of 50 liver cases are anonymized for this study. The liver contours are manually delineated in each case. Forty (40) cases are used for training, and ten (10) are reserved for validation. Dual-tree complex wavelet transform(DTCWT) are used in the deep learning U-net model. DTCWT provides multiresolution and sparse representation properties into the structure of an image. The deep convolution model provides the hierarchical structure to extract information from different feature levels. Similar to traditional U-Nets, we attempt to preserve information flow and details by interchanging maxpooling layers with discrete wavelet transformation and upsampling layers with inverse wavelet transformations.
Results: The segmentation mask of the liver is generated by feeding CT images into the wavelet U-Net. Evaluation of the wavelet U-Net image segmentation model is done with dice and pixel-wise accuracy. The model achieved a dice score of 0.896 and the pixel-wise accuracy of 0.83% in the validation dataset. This model is found to have achieved comparable results as currently available commercial packages. Further adjustments of the model parameters are in progress and can potentially improve its performance.
Conclusion: This Wavelet U-Net can currently be used as an automated quality assurance tool for commercial packages. Further development is in progress to extend its capability to segmenting liver targets.
Segmentation, Computer Vision, Convolution