mini-project-cnn

We are building a multi-class classification neural network aiming to classify tumor as benign (BEN), malignant (CAN) or normal (NOR).

Tech stack

listed libraries below were used to build the project.

streamlit: a quick way to build application interface for machine learning models
tensorflow and keras: for building and training the model

Usage

make a virtual environment activate it and install dependencies by running in app directory

python3 -m venv env && source .env/bin/acivate && pip3 install -r requirements.txt

run notebook tumor_classifier.ipynb in notebooks directory to generate model directory
run application with this command in app folder

streamlit run app.py

Model architecture definition

Convolutional Neural Network (CNN) is the suitable architecture for treating image datasets. Our choices are justified referring to this article

How many convolutional layers do we want ?

A CNN typically has three types of layers: a convolutional layer, a pooling layer, and a fully connected layer.

We built a sequential model with layers from the Keras library. Our model is built of:

Convolution layers (3)
- takes as parameters
  - input shape of data (only for the input layer)
  - the number of kernels
  - the size of kernels
  - activation function
- benefit
  - extract features out of the image using feature matrices
Pooling layers (3)
- takes as parameters
  - size of the pooling matrix
- benefit:
  - downsamples the output of a convolutional layers by sliding the filter of some size with some stride size and calculating the maximum (as we used a MaxPooling layer) of the input
Flatten layer (1)

Used to convert all the resultant 2-Dimensional arrays from pooled feature maps into a single long continuous linear vector so it later the data can be fed to a fully connected layer

Dropout layer (1)

Helps prevent overfitting by randomly setting input units to 0 with a frequency of rate at each step during training time

Dense layer (fully connected layer) (1) Neurons in this layer have full connectivity with all neurons in the preceding and succeeding layer.

Helps to map the representation between the input and the output. It can be computed as usual by a matrix multiplication followed by a bias effect

Output layer (1)

Gives the actual output of the probability of an image belonging to each of the three classes.

What should be the activation function of each layer ?

Negative weights affects badly the neural network, ReLu function eliminates negative weights by setting them to zero so we use it for all layers except for the output layer.

The output layer contains three nodes (one for each class). The softmax acivation function normalize the output of a network to a probability distribution over predicted output classes

How many hidden units should each layer have ?

Data is less complex is having fewer dimensions or features then neural networks we use lesser units for hidden layers with 16 to 64 kernels to convolutions and 128 node for hidden layer. The output layer must contain 3 and only three nodes as we have three classes.

Model training

We will define steps for training our neural network.

Dataset Structure

Link to dataset

In order to flow_from_directory method of the ImageDataGenerator data should be structured this way

|- ddsm
  |- train
    |- BEN
    |- CAN
    |- NOR
  |- validation
    |- BEN
    |- CAN
    |- NOR

We then construct two BatchDataset objects for train and validation directories. Each BatchDataset object contains normalized objects and one-hot-encoded labels.

Dealing with imbalanced data

Data is remarkably imbalanced so to prevent reducing data we penalized the model differently for each class depending on how much data we have for each one.

Feeding data

In order to accelerate the learning speed we used steps_per_epoch and validation_stepsparameters to treat data in batches. We also passed validation data to make sure we do not have overfitting. We set verbose=1 to visualize metrics for each epoch.

Model performance estimation

In order to estimate model performance, we did the following steps:

Plotted the epochs' history (loss_function and accuracy) for both training and validation datasets.
Plotted confusion matrix
Printed the overall accuracy
Printed classification report

aminbenmansour/mini-project-cnn