ImageClassification-MultilayerPerceptrons-CNNs

Contributors

Created by Marc-Antoine Nadeau, Jessie Kurtz, and Baicheng Peng

Overview

The complexity of modern medical imaging datasets has driven the need for advanced machine learning models capable of achieving accurate and efficient classification. This repository explores the OrganAMNIST dataset, a challenging benchmark derived from medical imagery, to evaluate the effectiveness of different neural network architectures for organ classification.

This project emphasizes the importance of hyperparameter optimization (e.g., learning rate, batch size, and regularization strength) and explores the trade-offs between computational efficiency and performance.

We investigate and compare the performance of:

Multilayer Perceptrons (MLPs):

Foundational models in machine learning, which struggle with capturing spatial dependencies in image data.

Convolutional Neural Networks (CNNs):

Specialized for spatial feature extraction, leveraging convolutional layers to create hierarchical representations.

Pre-trained MobileNetV2:

A lightweight architecture that transfers knowledge from large-scale datasets, achieving high accuracy with reduced training time.

File Structure

How to Run

Multilayer Perceptron

1. Navigate to the src/MLP folder.
2. Run the PrepareDataset.py file.

Convolutional Neural Network

1. Navigate to the src/CNN&MobileNet folder.
2. Run the CNN_28_128.py and cnn_hypara_exp.py files.

Resnet

1. Navigate to the src/CNN&MobileNet folder.
2. Run the MobileNet.py file.

Results

Any resulting figures/graphs will be saved in their respective folders:

• Results-MLP: Contains the results for the Multilayer Perceptron (MLP) model.
• Results-CNN: Contains the results for the Convolutional Neural Network (CNN) model.
• Results-ResNet: Contains the results for the ResNet model.

Key Features

Dataset:

• OrganAMNIST: A medical imaging dataset with grayscale images of 11 organs, resized to 28x28 pixels and 128x128 pixels for efficient experimentation.
• Dataset split includes training and test sets for rigorous evaluation.
• Normalization was done on the data using the Z-score technique seen in class.

Architectures Studied:

• MLPs: Baseline models that highlight the limitations of fully connected layers for image data.
• CNNs: Tested with various configurations (e.g., layer depth, kernel size, pooling strategies) to optimize performance.
• MobileNetV2: A pre-trained model offering high accuracy and efficiency.

Hyperparameter Optimization:

• Investigates learning rate, batch size, and regularization techniques (L1 and L2) to balance accuracy and computational cost.

Performance Metrics:

• Accuracy and Recall: For comparing model effectiveness.
• Training Time: For assessing computational efficiency.
• Robustness: Evaluating the model’s performance across different configurations.

Applications

This repository provides insights into designing neural networks for medical image classification tasks. The findings are particularly relevant for:

• Developing robust, efficient AI models for healthcare applications.
• Exploring transfer learning for improving model accuracy on small, specialized datasets.
• Balancing computational cost with performance in resource-constrained environments.

Results

Performance Overview

• MLPs: Struggled to achieve high accuracy due to the lack of spatial feature extraction capabilities.
• CNNs: With optimal configurations, achieved 0.8325 accuracy, highlighting the importance of convolutional layers and pooling strategies.
• MobileNetV2: Achieved the best accuracy of 0.9249, showcasing the benefits of transfer learning for medical image classification. Delivered a balance between accuracy and computational efficiency.

Prerequisites

Make sure you have the following Python libraries installed:

• numpy
• matplotlib
• tensorflow
• OrganAMNIST

You can install them using pip:

pip install numpy matplotlib tensorflow OrganAMNIST