Improving the Neural Segmentation of Blurry Serial SEM Images by Blind Deblurring

Figure 10

Illustration of consecutive EM images for the training of the 3D segmentation model. EM volume: original EM volume with focused and blur images and its 3D segmentation results; DeblurGAN-V2, SRN, and BANet: deblurring EM images according to the method from Kupyn et al. [9], Tao et al. [8], and Tsai et al. [18], respectively, and its 3D segmentation results; proposed: deblurring EM images according to the proposed deblurring method and our 3D segmentation results; label: annotated ground truth. Focused EM images are not replaced, such as the fifth and seventh columns. The rest of the deblurring images are implemented with different well-performed deblurring methods. Note that deblur-to-segmentation learning-based approaches, GAN-based or not, are better than the single 3D segmentation model with blur augmentation. Our proposed method has better performance on recovery of cell features and spatial noise distribution, leading to higher segmentation accuracy.