/SkinLesionAI

Notebooks of pre trained models using the HAM10000 dataset

Primary LanguageJupyter NotebookMIT LicenseMIT

SkinLesionAI

GitHub repository size

Google Colab Google Drive Jupyter Notebook Kaggle Python SQLite

Keras NumPy OpenCV Pandas ScikitLearn TensorFlow


Description

This repository refers to a final project of Computer Engineering major which contains CNN notebooks using the HAM10000 dataset, a dataset with 7 skin cancer classes:

  • Basal cell carcinoma;
  • Benign keratosis;
  • Bowens disease;
  • Dermatofibroma;
  • Melanocytic nevi;
  • Melanoma;
  • Vascular lesion.

Objective

Study the impact of configurations and techniques using vision transformer (accuracy and transfer learning only) and CNN models:

22.4 GB of model data were generated.

  • Mainly techniques used:

    • Data augmentation;
      • Image transformations;
      • Generative Adversarial Networks.
    • Segmentation;
    • Transfer learning.
  • Metrics extracted:

    • Accuracy;
    • Loss;
    • Sensibility (Recall);
    • Specificity;
    • F1-score;
    • AUC;
    • Precision;
    • Confusion Matrices.
      • Multiclass;
      • Per class.

Utilization of Different Models and Convolutional Neural Network Techniques to Skin Lesion Classification.pdf