Implementation of the networks used in the paper "Investigating Shift-Variance of Convolutional Neural Networks in Ultrasound Image Segmentation."

The vanilla U-Net, referred to as Baseline in the paper, where all four downsampling layers are conventional MaxPools (kernel=2×2, stride=2):

lpf_size = None
model_baseline = PBPUNet(n_channels=1, n_classes=1, lpf_size=lpf_size).to(device)

A BlurPooled U-Net, referred to as BlurPooling m x m in the paper, where all four downsampling layers are MaxBlurPools with anti-aliasing filters of the same size (m x m):

lpf_size = [3, 3, 3, 3] # Valid values for m are 2, 3, 4, 5, 6, and 7.
model_blurpooled_3 = PBPUNet(n_channels=1, n_classes=1, lpf_size=lpf_size).to(device)

The Pyramidal BlurPooled U-Net (PBPUNet), referred to as PBP in the paper, where the size of anti-aliasing filters gradually decreases at each downsampling layer from the first (shallow) to the fourth (deep) one. Filter sizes used in the paper for the PBP method: 7x7, 5x5, 3x3, 2x2. Other combinations can also be used, where valid filters sizes are 2, 3, 4, 5, 6, and 7.

lpf_size = [7, 5, 3, 2]
model_pbp = PBPUNet(n_channels=1, n_classes=1, lpf_size=lpf_size).to(device)

The list of three publicly available datasets employed in the paper:

• First Ultrasound Dataset: Breast Ultrasound Dataset B
• Second Ultrasound Dataset: Breast Ultrasound Images Dataset (Dataset BUSI)
• Brain MRI Dataset: LGG Segmentation Dataset

If you found this code useful, please cite the following papers:

• Sharifzadeh, M., Benali, H. and Rivaz, H., 2022. Investigating Shift-Variance of Convolutional Neural Networks in Ultrasound Image Segmentation. IEEE Transactions on Ultrasonics, Ferroelectrics, and Frequency Control.
• Sharifzadeh, M., Benali, H. and Rivaz, H., 2021, September. Shift-Invariant Segmentation in Breast Ultrasound Images. In 2021 IEEE International Ultrasonics Symposium (IUS) (pp. 1-4). IEEE.
• Zhang, R., 2019, May. Making convolutional networks shift-invariant again. In International conference on machine learning (pp. 7324-7334). PMLR.