Projects Goal: Classification of colorectal cancer histology.
Environment: Jupyter Notebook, Google Colab, Python
Building a model which classifies human colorectal cancer tissues into 8 different classes.
colorectal_histology- 5,000 150x150 images of human tissues with different textures from 8 classes with labels.colorectal_histology_large- 10 5000x5000 images of human tissues without labels.
First Dataset:
-
seperating images from the first dataset into 90% train and 10% test
-
using a VGG16 model with a custom classifier on top to fit images from first dataset.
-
freezing first layers of the VGG16 and training the classifier on the images.
-
data augmentation, unfreezing and fine tune to get better results.
Second Dataset:
-
cropping the large (5,000x5,000) images into 150x150 windows to make it fit the models input
-
predicting the class of each cropped window and combining the windows back to 5000x5000 image for each image from second dataset
Final result:
- multi-class visualization when each predicted class gets its own color
- heatmap of the probability of cancer (1 - high probability , 0 - low probability)
more info and code can be found in the notebook - CancerClassification.ipynb