Haoyu Lan1 and Farshid Sepehrband1
1USC Stevens Neuroimaging and Informatics Institute, University of Southern California, Los Angeles, CA, United States
1USC Stevens Neuroimaging and Informatics Institute, University of Southern California, Los Angeles, CA, United States
In this work, we proposed an unsupervised learning method for perivascular spaces segmentation by combing filter-based image processing technique and deep learning algorithm. The hybrid method improved the segmentation performance and eliminated the needed for manual annotation
Figure 1. The architecture of the proposed unsupervised learning segmentation method UPSS. UPSS is composed of mainly two parts: a Frangi filter as convolution neural network (CNN) with fixed gaussian kernels and a simple convolutional neural network like Unet. The results from these two parts are used as inputs of a conditional random filed (CRF) as the recurrent neural network (RNN) to perform segmentation post-processing. The three parallel backpropagations are conducted during each training step to effectively train all the weights and parameters of UPSS.
Figure2. Qualitative assessment of segmentation results of Unet and UPSS. The input image a., Frangi filter result b., UPSS result c., Unet result d., comparison between Frangi filter result and UPSS result e., comparison between Frangi filter result and Unet result f.