Awesome

2MSPK-Net

This repo is the official implementation of ['2MSPK-Net: A Nuclei Segmentation Network Based on Multi-Scale, Multi-Dimensional Attention, and SAM Prior Knowledge']

<p align="center"> <img src="https://github.com/ThirteenYue/2MSPK-Net/blob/master/prior.png" width="50%" height="50%" /> </p> We proposed a segmentation method based on SAM prior knowledge guidance strategy, and the above is a schematic diagram of integrating SAM prior knowledge.For detailed method introduction, please read the original article

Requirements

Install from the requirements.txt using:

pip install -r requirements.txt

Usage

Note: If you have some problems with the code, the issues may help.

1. Data Preparation

1.1. GlaS and MoNuSeg Datasets

🔥 The original data can be downloaded in following links:

MoNuSeg Dataset - Link (Original)
TNBC Dataset - Link (Original)

Then prepare the datasets in the following format for easy use of the code:

├── Dataset
    ├── MoNusg
    │   ├── test
    │   │   ├── boundary_priors
    │   │   ├── images
    │   │   ├── masks
    │   │   └── seg_priors
    │   ├── train
    │   │   ├── boundary_priors
    │   │   ├── images
    │   │   ├── masks
    │   │   └── seg_priors	
    │   └── val
    │   │   ├── boundary_priors
    │   │   ├── images
    │   │   ├── masks
            └── seg_priors

2. Training

During the training process, the data were uniformly resized to $256\times256$ pixels and data augmentation was applied, including affine transformation, random flipping, and random rotation. Gradient descent was performed using the Adam optimizer with $\beta_1$ set to 0.9 and $\beta_2$ set to 0.999. The initial learning rate was set to $1\times{10}^{-4}$, and an adaptive learning rate decay strategy was employed. If the loss on the validation set did not decrease after every 20 epochs, the learning rate was reduced by a factor of 0.5. The batch size was set to 4, and the training was completed after 600 epochs.

2.2 Pre-training

We didn't use any pre-trained weights

3. Testing

We also added SAM prior area maps and contour maps to the test data set

You can generate area maps and contour mapsarea and contour plots using SAM_prior We will announce other test and visualization codes later

Reference

UNet:https://github.com/LeeJunHyun/Image_Segmentation#u-net
UNet++: https://github.com/qubvel/segmentation_models.pytorch
U2Net:https://github.com/huangmozhilv/u2net_torch
Attention U-Net: https://github.com/bigmb/Unet-Segmentation-Pytorch-Nest-of-Unets
TransUNet: https://github.com/Beckschen/TransUNet
Swin-Unet: https://github.com/HuCaoFighting/Swin-Unet

Contact

Weight files and evaluation codes can be obtained by contacting the author ——Gongtao Yue(thirteen_yue@163.com)