DL - CNN Tensorflow

Vision - Brain Tumor Classifier & Live Prediction in Browser.

For this project, I built a Binary-Classifier to identify normal MRI brain scans vs with tumor MRI brain scans. This project aim was to achieve above 90% top-1 accuracy. I was able to get the model to predict the class of the scans with 97% accuracy with out using transfer learning.

To get these results I build a non-complex Convolutional Neural Network (CNN) model. This created time efficiencies and solid results. Further experiments are needed to improve the accuracy and overfitting like regularization, transfer-learning or getting more data.

More Details

I started this project with binary classification problem by making a model that was able to classify normal vs tumor brains. This experiment was done on a simple CNN models, having no more then 3 Conv2D (32, 32, 64 filters respectively) hidden layers, MaxPooling layer, dense layer, and dropout layer. These models were able to achieve above 95% accuracy. But from the loss and accuracy curve between train data and validation data, it was clear that model was overfitting because the distance between the train and val graph lines were large. Due to less limited time I stoped the project there and didn't regularize or fine-tuned the model. Other Details are in the following sections.

• Exporatory Data Analysis

  The experiment data was downloaded from Kaggle Br35H. It contained total 3000+ images, the data was small as compared to the data needed for ANN training. The dataset contained 3 folder, 'yes' -> with tumor, 'no' -> no tumor and 'pred'. So I used tensorflow data API to load and preprocess the data and split the dataset using sklearn train_test_split into 80-20 train and val datasets. The 'pred' images were used in the testing process.

  	Experiment #1
  Data Source	Kaggle
  Classes	2 (yes / no) tumor
  Individual Class	1500 x 2 = 3,000
  Train Data	2,388 Images
  Test Data	612 Images
  Normalization	Yes
  Data Augmentation	No
  Data loading	tf.data API

  Few examples from the traning dataset.

  
• Model Building

  The experiments was done on CNN architecture model, which was trainied 10 epochs (less epochs to avoid overfitting as it was trained only on 2,388 images)

  	Experiment #1
  Architecture	CNN
  Hidden layers	3 Conv2D, 3 MaxPooling, Flatten, Dense, Dropout
  Hidden Width	(32+ 32+ 64+ 64), dropout = 0.5
  Callbacks	None
  Epochs	10

• Model Performance

  Model achieved 97% accuracy which was not surprising as it was trained on less data and epochs, and it was overfitting, Nevertheless it was able to achieve the accuracy as of our target goal 90+% top-1 accuracy.

Predictions & live Prediction in Browser

For prediction I used the 'pred' images which were unseen for the model. These images were also preprocessed to match there dimension, shape and scale. Out of 9 model predicted all correct with high probability scores.

To view the source code, dataset, model, evaluation metrics (Accuracy), please visit my Github page.

Github page