Training a Deep Neural Network (DNN) to classify brain MRI scans for detecting Dementia versus Non-Dementia patients. This project aims to use a simpler architecture with fewer parameters while still achieving good accuracy. We focus on a data-centric approach, emphasizing data augmentation and fine-tuning hyperparameters rather than complex model architectures.
The dataset is sourced from Kaggle and can be found at Alzheimer's Dataset
The main objective of this project was not to create a model with the highest state-of-the-art (SOTA) accuracy but to understand how data affects model performance. By defining a simple CNN model and keeping it fixed, we focused on using data augmentation techniques and hyperparameter tuning to improve accuracy.
- PyTorch
- Python
- MLFlow
- Data Augmentation
- Hyperparameter Tuning
The dataset contains images for 4 classes:
- Mild Demented
- Moderate Demented
- Non Demented
- Very Mild Demented
The class distribution is as follows:
The dataset exhibits significant class imbalance. To address this, I employed and compared two data augmentation techniques:
-
SMOTE (Synthetic Minority Over-sampling Technique): Equalized the instance counts across all classes.
-
Manual Augmentation: Boosted the number of instances in underrepresented classes while retaining the overall class distribution. This approach maintained the relative differences in instance counts, offering a better representation of real-world conditions
Result: The model performs better when the class distribution closely mirrors real-life scenarios, providing a more accurate reflection of actual conditions.
Several models developed on this dataset leveraged pre-trained architectures such as RESNET [1], VGG, and others, achieving remarkable accuracies nearing 94%. Conversely, a different approach employed a more straightforward Convolutional Neural Network model, yielding a validation set accuracy of approximately 80% [2]. This project endeavors to craft a simplified DNN model, boasting fewer parameters than expansive architectures like RESNET, all while upholding high accuracy standards.