biodatlab / bacteria-detection

Deep Learning-Based Object Detection and Bacteria Morphological Feature Extraction for Antimicrobial Resistance Applications

Deep Learning-Based Object Detection and Bacteria Morphological Feature Extraction for Antimicrobial Resistance Applications

See our paper on IEEEXplore.

Antibiotics are the primary drug for treating various kinds of infections occurring from bacteria and microbes. They work mainly by blocking the vital pathway of those organisms and stopping them from multiplying. Previous research shows that we can predict the antibiotics used on bacteria by visualizing their morphology. Here, we present object detection for detecting bacteria and identifying the antibiotics used on them with their mophological features for example DNA intensity, contour area, and min areaRect.

There are 8 common classes we are interested including E.Coli bacteria treated with Ampicillin, Ciprofloxacin, Rifampicin, Tetracycline, Mecillinam, Kanamycin, Colistin, and Untreated.

Object detection

The current results of single models and ensemble models are as follows.

Model	Backbone	Head	Neck	mAP	mIOU	AP(50)	AP(75)	AP (medium)	AP (large)	Config	Checkpoint
YOLOv2	Darknet-19			0.053	0.192	0.015	0.048	0.102	0.140	config	ckpt
Faster R-CNN	ResNet-50			0.041	0.097	0.031	0.005	0.045	0.325	config	ckpt
Cascade R-CNN	Res2Net-101 + DCNv2	SABL	PAFPN + DyHead	0.652	0.800	0.808	0.762	0.677	0.692	config	ckpt
YOLOX-M	CSPDarknet	-	PAFPN	0.621	0.755	0.902	0.835	0.711	0.796	config	ckpt
Cascade R-CNN	Res2Net-50 + DCNvV2	SABL	PAFPN + DyHead	0.680	0.802	0.820	0.779	0.704	0.628	config	ckpt
Ensemble Model				0.753	0.699	0.863	0.796	0.717	0.675

Bacteria Feature Extraction

We apply the following approach to extract bacteria features:

• Feature Pyramid Network (FPN) for automatic color manipulation
• Deep MAC for instance segmentation from the bounding box given by object detection models
• Open-CV for feature extraction. We extract 19 morphological features using OpenCV to obtain the perimeter, area, length, and width of the cell membrane, DNA, and color intensity (minimum, maximum, mean, median, and standard deviation of the green and blue color channels)

We compare downstream antibiotic classification with SVM using features extracted from our model and CellProfiler.

	Bacteria detection model	CellProfiler
Mean F1-score	0.76	0.796

This difference is considered acceptable because the number of bacteria that the CellProfiler can detect is significantly lower than the model’s but higher in terms of quality since the CellProfiler can only detect complete bacteria cells.

Dataset

The current dataset contains 900 images: Ampicillin (100), Ciprofloxacin (100), Rifampicin (100), Tetracycline (100), Mecillinam(100), Kanamycin(100), Colistin(100), and Untreated (200). We annotated the RGB version in PNG format of the bacteria images which is easier to visualize using labelme. Images are in TIFF format that we use for actual model training and testing.

High intensity image	Low intensity image

Installation

Download and activate the environment

• Download the virtual environment from here
• Move the virtual environment to the bacteria-detection folder
• extract the virtual environment

cd bacteria-detection
conda activate ./path/to/virtual_environment

Download Pre-trained Weight

We release the pre-trained model weight for reproducibility purposes. You can download the weights of all models by DVC

dvc pull

Back-end: FastAPI

cd webapp/backend
uvicorn app:app --reload

Front-end

We use ReactJS as our frontend. To run an application, install NodeJS here and run the following

cd webapp/frontend/bacteria-app
npm start

Web application

Upload a bacteria file(s)

Perform prediction: bbox with class (left) and bbox with an index number for feature extraction table (right)

Table of Feature Extraction results

Inference without Web Appplication

cd webapp/backend/inference
python inference_ensemble.py --imgs_folder <Path/to/Image/Folder>

The outputs will be in the output folder and include

1. CSV file of all detection results
2. CSV file of all extracted features

Training

YOLOv2 model is trained using darknet
Other object detection models are trained using mmdetection

Citations

If you use our project or the our paper, please cite as

@INPROCEEDINGS{10322010,
  author={Chotayapa, Korrawiz and Leethamchayo, Thanyatorn and Chinnawong, Piraya and Samernate, Thanadon and Nonejuie, Poochit and Achakulvisut, Titipat},
  booktitle={2023 15th Biomedical Engineering International Conference (BMEiCON)}, 
  title={Deep Learning-Based Object Detection And Bacteria Morphological Feature Extraction For Antibiotic Mode Of Action Study}, 
  year={2023},
  volume={},
  number={},
  pages={1-5},
  doi={10.1109/BMEiCON60347.2023.10322010}}

ACKNOWLEDGMENT

We would like to thank Poochit Nonejuie Ph.D. and Mr. Thanadon Samernate from the Institute of Molecular Biosciences that inspired us and prepared the dataset for this study.