NC-Net for Nuclei Instance Segmentation

Official implementation of Nuclei Probability and Centroid Map Network (NC-Net) for nuclei instance segmentation in histology images. The paper can be read Here.

Setup Enviorment

For our experiments we have been using anaconda package manager and we have added environment.yml for creating the same environment that we have used. For setup of enviorment please run the following command:

conda env create -f environment.yml

Datasets

We have trained the model on three Datasets that are CoNSeP, PanNuke and Lizard. The links to the Dataset Download pages are:

• CoNSeP
• PanNuke
• Lizard

PWD
|--environment.yml 
|
|--checkpoints
|
|--run_logs
|
|--src
|
|  # please either create it or symlink it
|--data 
   |
   |--consep
   |  |--train
   |      |--001.npy
   |      |--002.npy
   |      |--*.npy
   |      |--n.npy
   |  |--test
   |      |--001.npy
   |      |--002.npy
   |      |--*.npy
   |      |--n.npy
   |
   |--pannuke
   |  |--train
   |      |--001.npy
   |      |--002.npy
   |      |--*.npy
   |      |--n.npy
   |  |--test
   |      |--001.npy
   |      |--002.npy
   |      |--*.npy
   |      |--n.npy
   |
   |--lizard
   |  |--train
   |      |--001.npy
   |      |--002.npy
   |      |--*.npy
   |      |--n.npy
   |  |--test
   |      |--001.npy
   |      |--002.npy
   |      |--*.npy
   |      |--n.npy
   |

Model Training

Before starting model training go through config.py for setting up various hyper-parameters that are esseintial for the model training. To start training please run the command below:

python train.py

Note: The training job currently only works with single GPU and does not support CPU training.

Pre-Trained Weights

The pre-trained model weights of NC-Net for CoNSeP, PanNuke and Lizard can be found here.

Model Inference

The model inference can done by running the command below:

python inference.py

Note: The training job currently only works with single GPU and does not support CPU training.

Model Evaluation

For evaluation we are using using Dice Score, Aggergarted Jaccard Index (AJI), Panoptic Quality (SQ), Detection (DQ) and Panoptic Quality (PQ). The model evaluation can be done using the command below:

python verify.py

Note: The training job currently only works with single GPU and does not support CPU training.