Different deep learning approaches in detecting various abnormalities via Chest X-Ray images

In this work, we would like to introduce 2 of the highlighted neural architectures in the field of object detection, which are: DEtection TRansformer and You Only Look Once. After that, we shall compare the performance between 2 these architectures via applying into the problem of detecting various abnormalities using CXR images. And finally, we will introduce our demo for this work.

Table of contents

1. Introduction
2. Repo structure
3. Demo
4. Experimental configuration
5. Pretrained model
6. Results
7. References

---

1. Introduction

😇 Our seminar towards this problem is here

Dataset

• Our dataset is taken from VinDr-CXR.

• Our dataset consists of 18 000 postero-anterior (PA) view Chest X-Rays (CXR) scans from a set of more than 100 000 raw images. They are all annotated by a group of 17 radiologists with at least 8 years of experience.

• Among 18 000 CXR scans, there are 5 000 scans served as training set, and 3 000 others as test set. The original size was 1024 x 1024 but we did resize them into 256 x 256 and change image format from DICOM into PNG.

Architectures

Here we use 2 architectures, which are: DEtection TRansformer and You Only Look Once.

• DEtection TRansformer (DETR): Proposed by Nicolas Carion and Francisco Massa from Facebook AI in 2020. DETR was first introduced to eliminate handmade's interference on postprocessing step effectively but still maintain the high performance compared to other methods.

• You Only Look Once (YOLOv5): Object recognition systems from the YOLO family are often used for vehicle recognition tasks, and have been shown to outperform other target recognition algorithms. YOLOv5 has proven to significantly improve the processing time of deeper networks. This attribute will gain in importance when moving forward with the project to bigger datasets and real-time detection.

2. Repo structure

• assets: images used for this readme
• Model
  • DETR.ipynb
  • YOLOv5.ipynb
• gitignore
• LICENSE

3. Demo

How to run model in local

Download nodeJS

• First, you need to download NODEJS. Node.js is an open-source, cross-platform, JavaScript runtime environment. To get more infomation about Node.js, please see the Node.js Website.

Install dependencies

• In your terminal, run this script to init:

chmod +x src/dependencies.sh && ./src/dependencies.sh

• Run

node src/index.js

Screenshot from our demo

4. Experimental configuration

• Proceeded in 30 epochs with GPU: NVIDIA @ Tesla P100-PCIE-16GB, RAM: 26GB and Pytorch framework
• Both architectures are trained and validated using 5-fold cross validation on training set; then tested on test set
• Use pretrained YOLOv5x and pretrained DETR on COCO val-2017
• Use ResNet50 as their own CNN backbone
• DETR’s learning rate is 3e-5 and YOLOv5x’s learning rate is 0.01

5. Pretrained model

• YOLOv5x
• DETR

6. Results

Private score on Kaggle's VinBigData Chest X-ray Abnormalities Detection contest

	Private score (mAP)
Rank 1	0.314
Rank 2	0.307
...	...
Rank 1005	0.137
Our YOLOv5	0.136
...	...
Rank 1099	0.064
Our DETR	0.062
...	...

Training curve and valid loss curve on 2 architectures

Comparison between 2 architectures

An example of output for a batch of 32 images

7. References

[1] Carion, N., Massa, F., Synnaeve, G., Usunier, N., Kirillov, A., & Zagoruyko, S. (2020, August). End-to-end object detection with transformers. In European Conference on Computer Vision (pp. 213-229). Springer, Cham.

[2] Stewart, R.J., Andriluka, M., Ng, A.Y.: End-to-end people detection in crowded scenes. In: CVPR (2015)

[3] Nguyen, H. Q., Lam, K., Le, L. T., Pham, H. H., Tran, D. Q., Nguyen, D. B., ... & Vu, V. (2020). VinDr-CXR: An open dataset of chest X-rays with radiologist's annotations. arXiv preprint arXiv:2012.15029

[4] G. Jocher, A. Stoken, J. Borovec, A. Chaurasia, L. Changyu, V. Abhiram, A. Hogan, A. Wang, J. Hajek, L. Diaconu, Y. Kwon, Y. Defretin, A. Lohia, B. Milanko, B. Fineran, D. Khromov, D. Yiwei and F. Ingham, ultralytics/yolov5: v5.0 - YOLOv5-P6 1280 models, AWS, Supervise.ly and YouTube integrations, Zenodo, 2021

[5] Cai, Z., & Vasconcelos, N. (2019). Cascade r-cnn: High quality object detection and instance segmentation. IEEE Transactions on Pattern Analysis and Machine Intelligence.

[6] Zhou, X., Wang, D., & Krähenbühl, P. (2019). Objects as points. arXiv preprint arXiv:1904.07850.

[7] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dollar, and C Lawrence ´ Zitnick. Microsoft coco: Common objects in context. In ECCV, 2014.

[8] Redmon, J., Divvala, S., Girshick, R. and Farhadi, A., 2016. You Only Look Once: Unified, Real-Time Object Detection.

[9] Redmon, J. and Farhadi, A., 2018. YOLOv3: An Incremental Improvement.

[10] Bochkovskiy, A., Wang, C. and Liao, H., 2020. YOLOv4: Optimal Speed and Accuracy of Object Detection