- Coded a Convolutional Neural Network using transfer learning to classify breast cancer images as benign or malignant.
- Used pretrained VGG 16 model and modified last layer.
- Added 2 fully connected layers in the end.
Used the breast cancer images provided in the BreakHis data set. It can be requested here: https://web.inf.ufpr.br/vri/databases/breast-cancer-histopathological-database-breakhis/
The original data folder looked like this
- numpy
- pandas
- matplotlib
- Opencv
- Keras
- Pytorch
I trained the data for different learning rates and number of epochs. I observed that with learning rate = 0.02 and number of epochs = 25 I get about 85% test accuracy as well as decreasing error graph. If I train it for 30 epochs, the model supoosedely overfits as seen in the test error. For learning rate = 0.001, the error does decrease, however, the accuracy is about 73% only.
Learning Rate | Number of Epochs | Training error | Testing error |
---|---|---|---|
0.02 | 25 | ||
0.02 | 30 | ||
0.001 | 25 | ||
0.01 | 25 | ||
0.05 | 25 |