The Complete Blood Count (CBC) Dataset has been used for automatic identification and counting of blood cells. Download the dataset, unzip and put the Training, Testing, and Validation folder in the working directory.
- Tensorflow-GPU==2.2.0 (tested on both 2.1.0 and 2.2.0)
conda install tensorflow-gpu - TF-slim==1.1.0
pip install tf-slim==1.1.0 - Weights:
downloadthe trained weights file for blood cell detection and put theweightsfolder in the working directory.
The code was originally written and developed with TensorFlow v1.x. The new updated version v2.0 included TensorFlow v2.x support, tested on both TensorFlow v2.1.0 and v2.2.0. You can download the previous version from here.
To detect the blood cells, simply run the detect.py file in the terminal or using an IDE. A step-by-step guideline of how to run the blood cell detection code in your computer is provided in this wiki.
If you have any trouble running the code and facing any error please feel free to create an issue or contact me.
A seven-step guideline of how to train on your own dataset is provided in this wiki.
The code was developed for the following blood cell detection paper. For a more detailed explanation of the proposed method, please go through the pdf of the paper. If you use this code or associated dataset, please cite this paper as:
Machine learning approach of automatic identification and counting of blood cells
@article{alam2019machine,
title={Machine learning approach of automatic identification and counting of blood cells},
author={Alam, Mohammad Mahmudul and Islam, Mohammad Tariqul},
journal={Healthcare Technology Letters},
volume={6},
number={4},
pages={103--108},
year={2019},
publisher={IET}
}
In some cases, our model predicts the same platelet twice. To solve this problem we propose a k-nearest neighbor (KNN) and intersection over union (IOU) based verification system where we find the nearest platelet of a selected platelet and calculate their overlap. We are allowing only 10% overlap between two platelets. If the overlap is more than that then it will be a spurious prediction and we will ignore the prediction.
| Before Verification | After Verification |
|---|---|
|
|
We have used our model to detect and count blood cells from high-resolution blood cell smear images. These test images are of the size of 3872 x 2592 way higher than the size of our trained images of 640 x 480. So, to match the cell size of our trained images we divide those images into grid cells and run prediction in each grid cell and then combine all the prediction results.
