We aim to identify malign and bening moles with a user-friendly prediction application. This is a slimmed-down version of the original repository.
We create a model from a pretrained model by re-training it with our image data. We then use a streamlit application to show predictions in a user-friendly interface.
This repo makes use of the HAM10000 dataset to train a model.
- make_binary_model.ipynb
- mole_detection_app.py
- Python 3.7.9
- CUDA 11.7 and CudNN
- Create and activate a virtual env
- Install the necessary libraries from the requirements.txt file in the main folder `
- To create a model, run all the files in the notebook
- To launch the app, type in the base folder
streamlit run mole_detection_app.py
A browser window will open which launches the app
- We will use the FastAI API to preprocess the data and build our model.
- We use the full HAM_10000 dataset containing 10057 images for training and validation of which 15% is reserved for the validation set
- Visual augmentation
We create 8 variations of any image (8 dihedral transformations) and add this to the dataset to increase the diversity of the image set - Data augmentation
We upsample with random crop and random resize
- We make use of the xceptionnet pretrained model
- We determine a good learning rate by choosing the valley in the learning rate curve
- We unfreeze the weights of the model and train it on the HAM10000 dataset in 24 epochs
- We validate the model by plotting the confusion matrix and the loss on both training and validation set.
The model shows the following results
| Metric | Outcome |
|---|---|
| True Positives | 12.23% |
| False Positives | 3.89% |
| True Negatives | 76.78% |
| False Negatives | 7.09% |
| Precision | 75.8% |
| Recall | 63.3% |
We worked on data preprocessing methods such as contrast, rebalancing and hair removal, which are not included nor used for the scope of this project.
- BeCode coaches (Louis & Chrysanthi)