Alzheimer's stage classification (Dr. Octupus Team)

Title

Alzheimer's Disease (AD) Patients' Brain MRI Stage Classification

Who

Introduction

Alzheimer's Disease (AD) is the most prevalent form of dementia, impacting millions and becoming increasingly problematic as the global population ages. Traditional diagnostic methods, such as medical history, neuropsychological testing, and MRI scans, are plagued with inconsistencies and inaccuracies that complicate both disease staging and patient care management. Accurate early diagnosis is critical as it enables timely clinical interventions, which can slow disease progression and significantly enhance the quality of life for patients. This is also the aim our team is working on. To address these challenges, we introduce a Convolutional Neural Network (CNN)-based model designed to improve the precision and recall of MRI-based diagnoses by tackling issues related to high model parameters and class imbalance. This approach is classified as a classification problem, where our model categorizes MRI images into various stages of Alzheimer's disease: Non-Demented (ND), Very Mild Dementia (VMD), Mild Dementia (MID), and Moderate Dementia (MOD). This stratification allows for more personalized patient care and targeted interventions at different stages of the disease.

Related Work

Our model is based on traditional CNN model with regularization techniques and a traditionally for-robust-regression loss. With the generalized formula described in the paper below, we were able to deduct a formula for multi-class classification using Huber loss. In our case, the huber loss is used to replace traditional categorical-crossentropy loss.

https://arxiv.org/pdf/2108.12627

Data

This dataset is curated to assist in the development of machine learning models for the classification of brain images by the severity of Alzheimer's disease. The dataset is divided into two subsets: Training and Testing, with approximately 6,500 images each. The images are labeled into four classes that represent the different stages of Alzheimer's severity.

The dataset is organized into two primary files:

Training: Contains the training images for model development.
Testing: Contains the testing images for evaluating model performance.

Each file includes the following classes:

MildDemented: Images representing mild stages of Alzheimer's.
VeryMildDemented: Images representing very mild stages of Alzheimer's.
NonDemented: Images with no signs of Alzheimer's.
ModerateDemented: Images representing moderate stages of Alzheimer's.

The dataset can be used to train algorithms for automated classification tasks, helping to identify the presence and severity of Alzheimer's disease from brain scans. It is crucial for the development of diagnostic tools that support early detection and intervention strategies.

Methodology

We employs a Convolutional Neural Network (CNN) with regularization and huber loss to classify MRI brain images into various stages of Alzheimer's disease. The CNN architecture, tailored for this purpose, processes through several layers to extract features and classify them as Mild Dementia (MID), Moderate Dementia (MOD), Non-Demented (ND), or Very Mild Dementia (VMD).

Input Layer: Normalized and enhanced MRI images are introduced to the model.
Convolutional Layers: Utilize filters to create feature maps, capturing essential characteristics from the input images.
Pooling Layers: Reduce spatial representation and computational load by selecting prominent features.
Dropout Layer: Implements neuron dropout during training to enhance model generalization.
Regularization Layer: Further regularize our model due to training and validation discrepancy, using kernel regularizer (Elastic Net) and activity regularizer (L2). -Softmax Layer: Flattened feature data are fed into dense layers, mimicking artificial neural networks and leading to a SoftMax layer for classification.
Loss Function: Employs a deduction from huber loss to evaluate the model's predictions and optimize performance.

The model's architecture is designed to mitigate overfitting while maintaining high efficiency in identifying Alzheimer's disease stages.

Metrics

The definition of "success" for our project is to beat the base line model's performance. While accuracy is an important metric, it may not fully capture the model's effectiveness in a clinical setting. Therefore, in addition to accuracy, we will employ a suite of more nuanced metrics to evaluate our model's performance thoroughly.

Precision and Recall: Critical for understanding the model's performance in the context of each specific class, where false positives and false negatives carry different implications for patient outcomes.

F1 Score: As the harmonic mean of precision and recall, the F1 score offers a single measure to balance these two metrics, which is crucial in the presence of class imbalance.

Ethics

What broader societal issues are relevant to your chosen problem space?
Why is Deep Learning a good approach to this problem?

In today’s life, doctors are often short-stuffed. Uneven distribution of medical resources has become one of the major issues we are facing. Our project intends automate classifying process of MRI images, specifically for AD patients, and consequently helping doctors to be more efficient.

Deep Learning is a good approach because with today’s advanced algorithm, image classification is much more accurate than before, and thus is suitable for initial assessment. We are hoping this could be an early step towards more advanced AI involvement in medical industry.

Admittedly, Purely AI-driven assessment is unrealistic in medical industry. We hope more automated classification and assessment could be used in an early assessment or an aid for the doctors to further interpret.

Division of labor

Qinmiao Deng: Project Lead and Model Optimization

Overseeing the project timeline, ensuring that milestones and deadlines are met efficiently.
Synthesizing team inputs and feedback to refine the model and its applications.
Ensuring that the project adheres to specified guidelines and objectives as laid out for the school project.

Kaishuo Zhang: Data Management and Preprocessing

Managing the acquisition and preprocessing of datasets to be used in training and testing our new model.
Performing initial data analysis and ensuring the quality of data inputs that feed into the model.
Maintaining data integrity and compliance with simulated clinical data use standards.

Kira Shen: Model Development and Technical Testing

Designing and developing the convolutional neural network architecture our model.
Implementing technical enhancements and optimizations such as dropout layers and various activation functions.
Documenting the development process and preparing a comprehensive report on the model's performance and technical specifications.

Xinbei Yu: Validation and Simulation Analysis

Developing the model architecture design and loss function design, using various metrics to measure accuracy, area under the curve (AUC), and other relevant performance indicators.
Compiling results and insights from model simulations to develop a final presentation for the school project.
Coordinating with the team to ensure consistency and reliability in the model's performance and create the final slidesfor it.