Assignmen3-AI

Project Assignment 3 Report Project Githul I. Introduction: In previous works on Alzheimer’s Detection stage classifications, many models such as Convolutional Neural Networks (CNN), Random Forests, Support Vector Machine (SVM), Naive Bayes, and K Nearest Neighbors were used as their MRI image classifiers. We would like to create a new model or improve upon an existing model. We will evaluate the results with performance measurements and compare the accuracy of our model with previous traditional models. The goal of our paper is to propose a new or modified model (modified CNN) that will improve the classification accuracy of Alzheimer’s dementia level using the same MRI image dataset. Il. Methodology (ideas we tried): We chose python as our programming language to develop and execute our machine learning models. Based on previous works, python was the most popular and easier programming language to use for building and running models. There are many machine learning libraries/packages available with python. Three of the most common machine learning libraries used in various literature reviews were PyTorch, TensorFlow, and Scikit-learn. We will primarily use the TensorFlow library for most of our implementation, since TensorFlow offers better visualization and easier deployment of our training models. TensorFlow is a good machine learning framework to support the building of neural networks like CNN. We will also incorporate other libraries for creating 2D plots with Matplotlib and to analyze the data with Pandas. The original dataset also used an Inception Model (Transfer Learning) which may be placing an artificial ceiling on accuracy. We plan on removing the Inception Model, and instead adding a layer of normalization and fine tuning for the model. Transfer Learning within our dataset: inception_model = Sequential() pretrained_model= tf.keras.applications.InceptionV3(include_top=False, input_shape=(224,224,3), pooling='avg',6classes=4, weights="imagenet') for layer in pretrained_model.layers: layer.trainable=False inception_model.add(pretrained_model)

Example of Fine Tuning in a similar dataset about Pneumonia: checkpoint_cb = tf.keras.callbacks.ModelCheckpoint("“xray_model.h5", save_best_only=True) early_stopping_cb = tf.keras.callbacks.EarlyStopping(patience=180, restore_best_weights=True) Another methodology we applied was with the InceptionV3 trained CNN. For this purpose we had to rescale the images to 176 x 208 pixel resolution and convert them to RGB format, so that the pretrained model could work effectively. The transfer learning method we are using requires certain layers of the CNN to be frozen to reuse pretrained model weights. The key steps in this transfer learning method are to freeze the base model weights, swap in a single dense layer for classification, implement early stopping based on validation AUC, and to train the model with InceptionV3. We then utilize the kerastuner package to optimize the hyperparameters. Dropout rates were set to .50 and .60 and the learning rate was set to .0001. The size of the convolution layer was set to 1024 as well as the dense layer. The final model layers are as follows: InceptionV3, dropout, batch normalization, conv2d twice, max pooling 2D, dropout, batch normalization, flatten, and two dense layers. The final iteration produced these loss statistics. Model AUC on Test Data:87.0% precision recall fl-score support mild 0.69 9.39 2.50 179 moderate 1.00 0.25 .40 12 normal 8.72 .77 8.74 640 very-mild 9.59 0.64 0.62 4438 accuracy @.67 1279 macro avg @.75 .51 .57 1279 weighted avg .67 0.67 .66 1279 Validation Loss by Epochs Validation AUC by Epochs : : y =P y =P Confusion Matrix 121 ® Taining loss 100 1 AUC:87.0% ~— Validation loss mid{ 7 0 51 58 400 0.95 1 < moderate{ 0 3 3 6 300 O 090 fi = v = normal { 12 0 153 200 0.85 1 ® Taining AUC very-mild { 20 0 140 100 080 ~ Validation AUC . ° -== Test AUC D> & R .§b T T T Q@ § ,@ 0 0 25 50 75 100 125 150 6& & & Epochs Predicted label Total Time:84.57 mins

Confusion Matrix MildDemented 32 ] 8 21 ModerateDemented 1 2 o 1 True Label 5 o 4 NonDemented 75 VeryMildDemented ° 25 Predicted Label Lastly, the current approach we’ll be using the Vggl18 model to train the data. The VGG-18 model is a type of deep neural network specifically designed for image recognition tasks. It's part of the VGG (Visual Geometry Group) series of models developed by researchers at Oxford University. What's notable about VGG-18 is its architecture: it consists of 18 layers, including convolutional layers, max-pooling layers, and fully connected layers. These layers work together to process and understand features within images. The following were the training and test loss results we received using the given model. w— Train - Validation 05 1 0 2 SO 75 100 125 150 175 200 Batches processed 0 P 50 75 100 125 150 175 Currently we are testing the model for benchmarks as well as working on augmenting the dataset.

lll. Data Loading and Processing: Our data source is a collection of MRl images downloaded from the Kaggle Alzheimer’s dataset to a local directory. We used TensorFlow’s tf.keras function to preprocess and load the images. The image set will be split up into 4 classes of dementia levels and the image size is set to 176 x 208. To visualize and verify the loaded images, we used Matplotlib to display the images in our model. We want to convert our model integer labels into one-hot encoded labels, since we are working with categorical data instead of noncontinuous data. The raw data does not come to us as feature vectors, so using the one-hot function will encode the text into a list of words for easier processing and evaluation of the data. IV. Building Our ML Model: CNN based approach was the most popular architecture in many literature reviews for Alzheimer’s Detection. This is due to CNN’s optimal performance in image classification, detection, and segmentation. We will be implementing a modified CNN approach model using TensorFlow’s Keras Sequential API to create and train our model. We will define a convolutional layer in Keras, which will take some input arguments, like specifying the shape. The CNN model will take tensors of shapes as an input. We will also need to define other layers such as the Rectified Linear Unit (ReLU), pooling, dropout, and fully connected layers. With the TensorFlow library, it does offer functions to support the building process of these layers. The number of filters in each convolutional layer will be something we need to experiment with and see the results. Adding many filters will allow for better learning capacity, but may lead to overfitting, which will cause issues with the training model. With our CNN-based model, we will modify the multi-layer convolutional network and try to improve upon the traditional approach using TensorFlow's machine learning library, by using a much larger and nuanced dataset. Additionally, the original dataset for Alzheimer's classification was imbalanced, meaning that the majority of the images (72%) were classified as demented, and only 28% were classified as non-demented Our modified approach will apply a layer of data balancing by accounting for the ratio of demented to non-demented images through the use of class weights. We also will be applying a layer of Fine Tuning. We want to fine tune both our model and our learning rate. Too high a learning rate will cause our model to diverge and too low a rate will cause our model to be slow. V. Model Evaluation: Model accuracy and model loss will be evaluated after the model training. We will compare its performance on both the training and testing datasets by plotting metrics in a visual graph. To achieve a more accurate assessment of our model, we will use different amounts of epoch cycle to compare the results. First we wanted to view the model loss and accuracy of the original dataset consisting of ~5,000 images. Then we want to view the same graphs for the augmented Alzheimer's dataset consisting of 40,000 images before and after removing the inception model. Original model accuracy & loss:

Paper Accuracy Brain MRI Analysis for Alzheimer’s Disease Diagnosis Using CNN-Based Feature Extraction and Machine Learning CNN softmax accuracy 96% Deep learning-based classification of healthy aging controls, mild cognitive impairment and Alzheimer’s disease using fusion of MRI-PET imaging CNN Adam optimizer accuracy 95.06% An Alzheimer’s disease classification method using fusion of features from brain Magnetic Resonance Image transforms and deep convolutional networks CNN with SGDM optimizer accuracy 85.4% Classification of cognitively normal controls, mild cognitive impairment and Alzheimer’s disease using transfer learning approach CNN with Structural transfer with thresholding and BRF classifier accuracy 79% Integrating convolutional neural networks, kNN, and Bayesian optimization for efficient diagnosis of Alzheimer's disease in magnetic resonance images CNN model with KNN classifier and Bayesian optimizer accuracy 94.96% Alzheimer's classification(Project) — Approach 1 Efficient B3 CNN model with 75% accuracy Alzheimer's classification(Project) — Approach 2 InceptionV3 CNN model with 91 % accuracy Alzheimer's classification(Project) — Traditional Sequential CNN model with 75 % accuracy(dropped to 60 with higher epochs) Alzheimer's classification(Project) — Current Approach Vggl18 CNN model with 95 % accuracy