devadethanr/alz_model

Alzheimer's Disease MRI Image Classifier Report

• 1. Introduction The goal of this project was to develop a deep learning model capable of classifying MRI images as indicative of Alzheimer's disease or a healthy brain. This report details the approach, methodology, and results of the project.
• 2. Dataset The project utilized the "Alzheimer_MRI" dataset hosted on Hugging Face Datasets. This dataset contains MRI scans of brains labeled with different stages of Alzheimer's disease. The images are in grayscale and come preprocessed in PIL (Python Imaging Library) format. Link : dataset Columns : images, labels Rows : 5.1k Labels : {0: 'Mild_Demented', 1: 'Moderate_Demented', 2: 'Non_Demented', 3: 'Very_Mild_Demented'}
• 3. Model Selection and Architecture Given the constraints of the Hugging Face free tier, the primary consideration was choosing a lightweight and efficient model architecture. The following options were evaluated: MobileNetV2: Known for its exceptional speed and small model size. EfficientNet-B0: A good balance between accuracy and efficiency. (5.3 million params) EfficientNet-B1: A good balance between accuracy and efficiency. (7.8 million params) SqueezeNet: Very small but less accurate than the other options. Ultimately, EfficientNet-B0 was chosen due to its balance of accuracy and efficiency. While slightly larger than MobileNetV2, it was expected to fit within the free tier's resource limits and potentially provide better performance for this medical imaging task. Link : Efficientnet-b0

• Hyperparameter Choices:

• • per_device_train_batch_size=32
• • num_train_epochs=25
• • fp16=True : Mixed precision training is enabled to leverage the GPU's Tensor Cores (if available) and speed up training.
• • load_best_model_at_end=True : This ensures that we use the best performing model
• • push_to_hub=True & hub_model_id & hub_strategy = "every_save"

• #. Results and Evaluation

The final model achieved : Training:

Validation: