Classification of polyps in colonoscopy images using Fastai deep learning
In 2018, I published a github repository CNN4Polyps about colonoscopy polyps detection (classification + localization into an image) with CNNs (https://github.com/muntisa/Colonoscopy-polyps-detection-with-CNNs) using Keras. I demonstrated that simple CNNs or VGG16 transfer learning could be used in only few minutes (on a GPU) to create good classifiers able to detect a polyp in colonoscopy images.
Reading twitts these days, I found a great presentation of Jeremy Howard about "Future of Individualized Medicine 2019" (https://twitter.com/jeremyphoward/status/1112810731773190144). During the presentation, I was suprized to find that on Modelzoo there is no shared pre-trained network even if there are a lot of public medical imaging dataset and easy tools for DL such as Fastai & Colab. Therefore, I decide to spend one hour to adapt the Fast.ai Lesson 1 notebook (https://course.fast.ai/videos/?lesson=1) for my dataset. To be faster, I used Google Colab (https://colab.research.google.com) with GPU support.
The current dataset was generated with the previous project CNN4Polyps starting from a public datase: 910 images for training and 302 images for validation. All the models are saved in the project root, so you can easy load them and test your data.
The current script demonstrated the ability to create a very accurate classifier for medical imaging with an accuracy of 0.99 using resnet50 transfer learning fine tuning. Due to the github file dimension limitation, there are open links to the saved networks in my public gdrive.
- Predict colon polyps in one image from test folder: Fastai-Colon-Polyps_Predict.ipynb.
- Re-train your better model using the current dataset: Fastai-Colon-Polyps.ipynb.
The best model was implemented as a web app using Flask: https://github.com/muntisa/Colon-Polyps-Fastai-App.
You can directly use the already created docker from https://hub.docker.com/repository/docker/muntisa/colon-polyps-fastai.
Run the local docker with the following command:
docker run -p 5000:5000 colon-polyps-fastai
Hf with DL! @muntisa