Skin Cancer Imaging AI Diagnostic Tool

Contributors

• Kruel, Kimberly
• McClure, Mistie
• Snyder, Todd

Program Structure

└───root
    └───datasets
        │   test_benign.pkl
        │   test_malignant.pkl
        │   test.pkl
        │   train_benign.pkl
        │   train_malignant.pkl
        └───train.pkl
    └───models
        │   model_adam_5.h5
        │   model_adam_scaled.h5
        └───model.h5
    └───utils
        └───conda_list.txt
    │   .env
    │   .gitignore
    │   main.ipynb
    │   dataset_creation.ipynb
    │   dataset_load.ipynb
    │   dataset_read.ipynb
    │   gradio.ipynb
    │   modeling_hypertuning.ipynb
    └───README.md

Data Processing

Step 1:

Images were downloaded from Kaggle to the local machine. The dataset_creation.ipynb file was run to create pickle files from the 4 datasets: test/benign, test/malignant, train/benign, and train/malignant. The pickle files were saved to /datasets. Due to size limitations, the pickle files were excluded from GitHub.

Note: This step does not need to be run to recreate/run other files.

Step 2:

Pickle files were loaded to AWS S3 storage. The dataset_load.ipynb file loads the pickle files from AWS and saves them locally.

Note: .env file is required to access AWS. To save costs, the files are saved into the local folder for developer use. This file only needs to be run once to create local files.

Step 3:

Pickle files are read from local. The dataset_read.ipynb reads the pickle files from local directory files that were created in Step 2.

Step 4:

Run select panels in modeling_hypertuning.ipynb. This file reads the pickle files, consolidates the files into train and test dataframes, and shuffles the data. Data is separated into X and y variables, training and test variables. Grid search was completed to determine best parameters. Models were created and saved to /models.

Note: Only select panels need to be run to recreate models.

Step 5:

Run gradio.ipynb to create and launch the Gradio app. This file reads the saved model and creates a Gradio app. The app allows the user to select their language, upload an image, and submit it. The user then receives an image prediction in their selected language.

Gradio Interface

The Gradio interface provides an easy-to-use web interface for the Skin Cancer Imaging AI Diagnostic Tool. This allows users to upload an image of a skin lesion and receive a diagnostic prediction.

Setup Gradio Interface:

1. Import Libraries
2. Load the Model
3. Create the Gradio Interface using gr.Blocks

For detailed code and methods used, please refer to the gradio.ipynb file.

To launch the Gradio interface, run the gradio.ipynb notebook. The main components of the interface are:

• Language Selection: Users can select the prediction language from a dropdown menu. The default language is English.
• Image Upload: Users can upload an image of a skin lesion.
• Submit Button: Users can submit the image for prediction.
• Prediction Output: The diagnostic prediction is displayed in a textbox.
• Questions/Chat Input: Open field to type out questions.
• Submit Question: Users can submit questions for answers from a chatbot.
• Answers to Questions Output: Answers to questions are provided by the chatbot.
• Clear Button: Users can clear the inputs and outputs.

Application Function

1. Launch Gradio Interface:

   jupyter notebook gradio.ipynb

2. Follow the Instructions:
   • Select the prediction language.
   • Upload an image of the skin lesion.
   • Click the submit button to receive the diagnostic prediction.
   • Use the clear button to reset the interface for a new prediction.

Screenshot

The interface will look something like this:

Resources

• utils/conda_list.txt contains the conda environment that this app was processed in.

External Link

• Hugging Face Space

Datasets

• Melanoma Cancer Image Dataset

License

• CC0: Public Domain

Sources

"The melanoma cancer dataset was sourced through a comprehensive approach to ensure diversity and representativeness."

Collection Methodology

"The images were collected from publicly available resources."

Notice

The AI tool has been developed for entertainment purposes and is not intended to provide medical advice.