RobinHCK/Real-time-detection-for-renal-pathology

A Keras implementation of YOLOv3 (Tensorflow backend)

Real-time detection for renal pathology

Table of Contents

• Presentation of the Project
• Prerequisite
• Workflow
  • Dataset
  • Train
  • Test
  • Perform detections on video
• Contact
• Acknowledgements

Presentation Of The Project

This research is carried out as part of the project Sys-MIFTA and has been been published in the conference IEEE 33rd International Symposium on Computer Based Medical Systems (CBMS). The paper is available here. This work is a deep learning project applied to medical images. The dataset contains WSI of stained H&E renal nephrectomies that are used by the YOLOv3 network to create a model capable of detecting glomeruli in real-time.

Prerequisite

• Before executing the sctipts, make sure you have correctly entered the configuration file: config.cfg
• The medical images used for this project are private data that we cannot share. You will need to use your own data. Here is the hierarchy that is expected by the workflow to work properly:

Workflow

1. Dataset

• Perform data augmentation:
  • python 1_augmentDataset.py --config config.cfg

2. Train

• Download the model pretrained on COCO dataset:
  • wget https://pjreddie.com/media/files/yolov3.weights
• Convert the weights for Keras:
  • python 2a_convertWeights.py yolo3/yolov3.cfg yolov3.weights model_data/yolo.h5
• Train the network:
  • python 2b_train.py

See the configuration file to know the model location

3. Test

• Create a file fill with detections performed by the network:
  • python 3a_test.py
• Compute and write metrics in datas.xlsx:
  • python 3b_computeMetricsOnTest.py --config config.cfg
• Draw detections on WSI with the best F1Score per scale:
  • python 3c_drawBestWSI.py --config config.cfg
• Draw graphics thanks to datas.xlsx:
  • python 3d_drawGraphics.py

Do not forget to test the right model, see the configuration file to know the model location

4. Perform Detections On Video

• python 4_yolo_video.py --input video/your_video.mp4 --output video/your_video_with_detections.mp4 --model model_data/yolo.h5 Do not forget to test the right model, see the configuration file to know the model location

Contact

Robin Heckenauer - robin.heckenauer@gmail.com

Acknowledgements

• ERACoSysMed project "SysMIFTA", co-funded by EU H2020 and the national funding agencies German Ministry of Education and Research (BMBF) project management PTJ (FKZ: 031L-0085A), and Agence National de la Recherche (ANR), project number ANR-15-CMED-0004.
• The High Performance Computing center of the University of Strasbourg. The computing resources were funded by the Equipex Equip@Meso project (Programme Investissements d'Avenir) and the CPER Alsacalcul/Big Data.