/Colorectal-Cancer-Histology-Transfer-Learning-and-Fine-Tuning

Transfer learning & fine-tuning in Tensorflow for classification of textures in colorectal cancer histology

Primary LanguageJupyter Notebook

Colorectal-Cancer-Histology-Transfer-Learning-and-Fine-Tuning

  • Fine-tuning an image classification model on a Colorectal-Cancer-Histology dataset.
  • This is a biological 8-class classification problem.
  • The dataset consists of 5000 images.
  • Each example is a 150 x 150 x 3 RGB image of one of 8 classes.
  • The class labels are: ("tumor", "stroma", "complex", "lympho", "debris", "mucosa", "adipose", "empty")
  • The state-of-the-art CNN, ResNet50V2, is used as base model.
  • The last 10 layers of the base model are unfreezed for fine-tuning.
  • Data augmentation is implemented for regularization.
  • Learning rate reduction callback is implemented.
  • F1-score and confusion matrix are visualized.
  • Accuracy of 94% is achieved on validation and test datasets.
  • Dataset Source: https://www.tensorflow.org/datasets/catalog/colorectal_histology
  • Dataset homepage: https://zenodo.org/record/53169#.XGZemKwzbmG