skshashankkumar41/malaria-detection-using-faster-rcnn

Given images of blood samples predict whether a patient has malaria or not using Tensorflow Object Detection API and Transfer Learning

Malaria Detection using Faster-RCNN and Transfer Learning

Given images of blood samples predict whether a patient has malaria or not using Tensorflow Object Detection API and Transfer Learning

Blood Sample Image	After Detection Image

Data

Data came from three different labs’ ex vivo samples of P. vivax infected patients in Manaus, Brazil, and Thailand. The Manaus and Thailand data were used for training and validation while the Brazil data were left out as our test set. Blood smears were stained with Giemsa reagent, which attaches to DNA and allow experts to inspect infected cells and determine their stage.

It can be downloaded from Kaggle

Approach

As the data is highly imbalanced and almost 98% cells are Red Blood cells, I made this problem a two stage classification task.

First Stage

Label all the cells other than RBC as Non RBC cells and then create a object detector that only detects either cell is RBC or NON RBC. For this task I used Tensorflow Object Detection API to fine-tune and train a Faster RCNN model pre-trained on coco dataset.

Second Stage

For all the NON RBC cells create a CNN Classifier to classify given a NON RBC cell which category it belongs (trophozoite, schizont, ring, leukocyte, gametocyte). I fine-tuned VGG-16 Model trained on imagenet data.

Two Stage Classification

So by combining the first stage and second stage we build a complete two stage classification model in which we will supply a blood sample image, Faster RCNN detector will detect and gives bouding boxes of RBC and NON-RBC in the image and then for all the NON-RBC we will crop the NON-RBC and pass that cell image to our VGG-16 classifier and it will predict the categories of NON-RBC.

Results

Training Data

MAP Scores

Counts

	Grount Truth Count	Model Truth Count
red blood cell	77420	71040
trophozoite	1473	1513
ring	353	311
schizont	179	178
gametocyte	144	157
leukocyte	103	87

Testing Data

MAP Scores

Counts

	Grount Truth Count	Model Truth Count
red blood cell	5614	4551
trophozoite	111	146
ring	169	95
schizont	11	6
gametocyte	12	7
leukocyte	0	0

Usage

Clone the repository and run the following commands from the terminal.

Setup Tensorflow Object Detection API

Go through this blog to setup TFOD on your system.

Install project dependencies from requirements.txt

 pip install -r requirements.txt

Creating .records and classes files

 python TFODRecordCreator.py

Training Faster RCNN

In the experiment folder paste the Faster RCNN resnet101 coco model in training subfolder and then run the following command from TFOD folder

python object_detection/model_main.py --pipeline_config_path faster_rcnn_malaria.config --model_dir experiment/training --num_train_steps 20000 --sample_1_of_n_eval_examples 10 --alsologtostderr

Export Faster RCNN Model

Run the following command from TFOD folder

python object_detection/export_inference_graph.py --input_type image_tensor --pipeline_config_path faster_rcnn_malaria.config --trained_checkpoint_prefix experiments/training/model.ckpt-20000 --output_directory output/models

Crop and Extract NON-RBC Images

 python CellExtractor.py  

Train VGG-16 on Extracted NON-RBC Images

python VGG16Trainer.py

Inference

python inference.py --imagePath "test.png"

Files

preprocessing.py
As data is present in json file this file creates annotated csv files which will be required by our Tensorflow Object Detection API.

TFAnnotation.py
This class defines the format of data required by Tensorflow Object Detection API.

TFODRecordCreator.py
This script creates the .record and classes files using TFAnnotation class which is required by Tensorflow Object Detection API to train Faster RCNN model.

plotting.py
Script for plotting of bouding boxes of RBC, NON RBC on images.

CellExtractor.py
Script for cropping and extracting NON-RBC images using bounding boxes from main image for the training of VGG-16 model.

VGG16Trainer.py
Script for training of cropped NON-RBC images using VGG-16 pretrained model.

trainDataInference.py, testDataInference
Script for evaluation of training and testing images using two-staged classification.

inference.py
Function to predict the result of blood sample image and also shows the image of input blood sample with labels of each cell and bouding box.

Contributing

Pull requests are welcome. For major changes, please open an issue first to discuss what you would like to change.