Fully Convolutional Networks for Instance Segmentation

11th place solution to Multi-organ nuclei segmentation and classification challenge (MoNuSAC 2020) organised at ISBI 2020.

The final submission scored an Average Panoptic Quality(PQ) of 0.1838 on the test set. More details about the challenge and other method can be found at URL.

`

Overview

In this work, we implement an end-to-end deep learning framework for automatic nuclei segmentation and classification from H&E stained whole slide images (WSI) of multiple organs (breast, kidney, lung and prostate). The proposed approach, called PatchEUNet, leverages a fully convolutional neural network of the U-Net family by replacing the encoder of the U-Net model with an EfficientNet architecture with weights initialized from ImageNet.

Since there is a large scale variance in the whole slide images of the MoNuSAC 2020 Challenge, we propose to use a patchwise training scheme to mitigate the problems of multiple scales and limited training data. For the class imbalance problem, we design an objective function defined as a weighted sum of a focal loss and Jaccard distance, resulting in significantly improved performance. During inference, we apply the median filter on the predicted masks in an effort to refine the segmented outputs. Finally for each class mask, we apply watershed algorithm to get the class instances.

Some preliminary results

Results reported in this table are from the validation set made from the entire training set provided for the challenge. These are results of multi-class segmentation of the 4 cell types.

Config	IoU(%)	Dice(%)
U-Net	77	78
Proposed	84	87

Requirements

• Python: 3.6
• Tensorflow: 2.0.0
• Keras: 2.3.1
• segmentation_models.
• OpenCV

Environment installations

Run this command to make environment

conda env create -f environment.yml

OR you can make you own environment by:

conda create -n yourenvname python=3.6 anaconda

Then install the packages

conda install -c anaconda tensorflow-gpu=2.0.0
conda install -c conda-forge keras
conda install -c conda-forge opencv
conda install -c conda-forge tqdm

The run conda activate yourenvname.

NOTE: segmentation_models does not have conda distribution. You can install by running pip install -U --pre segmentation-models --user inside your environment. More details at segmentation_models.

Dataset versions

• MoNuSAC_images_and_annotations : original dataset which has patient's whole slide images (WSI) and ground truth. (Given by challenge organizers)
• MoNuSAC_masks : contains binary masks generated from get_mask.ipynb.
• Testing Images: contains test images, without annotations. (Given by challenge organizers)
• data_processedv0 : contains all raw images and the ground truth masks.
• data_processedv4 : sliding window patchwise data from original images and masks in data_processedv0.
• data_processedv5 : 80/20 trainval split from data_processedv4.

Getting started

0. Clone repository (obviously!)
1. Make dataset folder
2. Put MoNuSAC_images_and_annotations in dataset folder
3. Run 0_get_masks.ipynb. You should get the MoNuSAC_masks folder in dataset
4. Run 1_data_process_MoNuSAC.ipynb to get raw images and their ground truth masks in data_processedv0.
5. Run 2_extract_patches.ipynb to get patches of images and gt masks from the previous raw version to get data_processedv4 and the 80/20 split data_processedv5.
6. Run 3_train.iynb. It trains PatchEUNet on data_processedv5.
7. Put Testing Images in dataset folder.
8. Run 3c_load_test_data.ipynb. Outputs test_images folder with all test images.
9. Run 4_inference.ipynb to get final prediction masks from test_images. (For visualization)
10. Run 4b_inference.ipynb to get final prediction masks according to MoNuSAC submission format.

Model

Weight file for the multiclass segmentation model is available at releases.

Acknowledgements

Model definition and baseline training scripts are based on https://segmentation-models.readthedocs.io/en/latest/install.html.

License

Your driver's license.