Experiment on Breast Cancer Classification

About

Deep learning application developed using Python and Keras with the aim of diagnosing breast cancer

Description

Based on the results and the network presented in this paper, I conducted an experiment on breast cancer classification.
I trained different KNN classifiers and deep models using the BreakHis dataset.

To achieve this, I used well-known deep-learning frameworks for Python such as:

• TensorFlow
• Keras
• scikit-learn

The goal of this work is to compare the performance of 4 different deep models that should perform well with images:
BreastNet, VGG16, VGG19, Inception-ResNet.
In details, the Inception-ResNet shares some observations made to build the network IRRCNN presented by the main paper. While in the last three cases the neural network implementation is directly provided by Keras, the former was presented in 2019, through the following paper, and was implemented in python by the same researchers (see breastnet.py).

Because of the small number of images contained in the dataset (7909 images), I used data augmentation, that is a regularization method, before feeding the networks.
Also, as optimizer I used SGD with 0.9 momentum and decay, since from the papers I read it seems to be a well-performing optimizer and kind of a standard for this type of data.

Sections:

1. Preprocessing

2. KNN classification

3. Data augmentation

4. Deep models

   1. Hyperparameters
   2. BreastNet
   3. VGG16
   4. VGG19
   5. Inception-ResNet

5. Comparison of results

6. Conclusion and further experiments

How to run

In order to replicate and to extend this experiment you need to:

1. download the BreakHis dataset
2. extract it in a folder named data
3. build the dev container using the Docker extension of VScode
4. run the notebook