LightNuclei: Simultaneous Nuclei Counting and Segmentation by a Light-weighted Deep Learning Model

This repository contains codes for an automated nuclei detection pipeline that is based on Yuanfang's gold medal solution in the 2018 International Data Science Bowl.

Please contact (gyuanfan@umich.edu) if you have any questions or suggestions.

Installation

Git clone LightNuclei:

git clone https://github.com/GuanLab/LightNuclei.git

Dependency

• python (3.6.5)
• tensorflow (or tensorflow-gpu) (1.9.0)
• keras (2.6.1)

Examples

The complete dataset for training can be downloaded from 2018 International Data Science Bowl.

Please download stage1_train.zip into data directory, and decompress there.

• Step 1. Preprocessing and split the data set:

  python LightNuclei.py

This step will generate a set of list files that split 80% of the images for training and 20% for testing (list_test). Furthermore, The training set is randomly split into 80% nested training (list_train_x.1) and 20% validation set(list_train_x.2) for five times.

• Step 2. Training stage:

  python train_nuclei.py

• Step 3. Prediction on test set:

  python test_nuclei.py /path/to/weight

/path/to/weight : .h5 weight file is passed to the program. Example is ./weights/pretrained.h5, or is produced and stored in ./logs/ after training the model in step 2.

This step will generate a set of folders named as vis_0, vis_0a, vis_1, vis_1a, vis_2, vis_2a, vis_3, vis_3a. Each folder contains a set of images that is a rotation/flip variant of the original test image set. The prediction is visualized in these images as binary mask.

• Step 4. Assemble the rotation/flip variants to make final prediction

  python assemble.py

  This step will generate a file named 'prediction.csv' in which each entry corresponds to an instance mask that is run-length encoded. Formatting details can be found here

• Step 5. Evaluation

  python eval.py

  This step outputs the mAP for each image in 'eval.csv'.

Example prediction is shown below.