Hippocampus Segmentation for Alzheimer's Induced Dementia Prediction from MRI Images

This repository contains the implementation of a project focused on segmenting the hippocampus from MRI images to predict Alzheimer's-induced dementia. By leveraging advanced segmentation techniques and improving existing architectures, this project aims to enhance the accuracy of hippocampus segmentation, an important step in understanding Alzheimer's disease progression.

---

Table of Contents

• Introduction
• Dataset
• Model Architecture
• Modifications and Enhancements
• Performance Metrics
• Resources Used
• Future Work
• Acknowledgements

---

Introduction

Alzheimer's disease is a progressive neurodegenerative disorder, and early diagnosis can help improve management strategies. Hippocampus segmentation from MRI images plays a critical role in assessing the extent of dementia. This project proposes improvements to segmentation architectures to achieve higher performance in this domain.

---

Dataset

Dataset used: MRI Hippocampus Segmentation

The dataset consists of:

• MRI images of the brain
  • 18900 MRI images of 100 patients in the training set
  • 6615 MRI images of 35 patients in the test set
• Corresponding segmentation masks for the hippocampus.

The dataset was preprocessed to normalize pixel intensities and resize images for input into the neural network. (128 x 128 x 3)

---

Model Architecture

The architecture for segmentation was inspired by existing state-of-the-art convolutional neural network (CNN)-based designs for image segmentation. Key layers include:

• Encoder-Decoder structure.
• U-Net-like skip connections for feature preservation.
• Batch normalization to stabilize training.

Base models and architectures were taken from:

• MultiResUNet
• Modified U-Net

---

Modifications and Enhancements

Several architectural modifications were introduced to improve segmentation performance:

1. Swish Activation Function:

   • Replaced ReLU with Swish for smoother gradients and better convergence.

2. SpatialDropout2D:

   • Added spatial dropout to prevent overfitting and improve generalization.

3. Bilinear Interpolation:

   • Used bilinear interpolation for upsampling, ensuring smoother feature maps during decoding.

4. Combined Loss Function:

   • Designed a combined loss function incorporating Binary Cross-Entropy and Dice Loss to optimize pixel-wise accuracy and segmentation overlap:
     • Binary Cross-Entropy ensures precise classification.
     • Dice Loss emphasizes accurate segmentation by accounting for overlap between predicted and true masks.

---

Performance Metrics

The modifications resulted in significant performance improvements, particularly in:

• Dice Coefficient: Measures the overlap between predicted and ground truth masks.
• Jaccard Index: Evaluates similarity between predicted and true segmentation regions.

Model	Dice Coefficient	Jaccard Index
MultiResUNet	0.897474389	0.87823319367
Modified U-Net	0.8914032548	0.872448485
Our Model	0.9057411749	0.88360496

---

Resources Used

• Platform: Kaggle
• RAM: 29 GB
• GPU: Tesla P100
• VRAM: 16 GB

Kaggle Notebook Link

---

Future Work

• Use the predicted masks for AD classification tasks.
• Extend the model to multi-class segmentation tasks.
• Experiment with other activation functions and loss functions.
• Integrate clinical data to improve prediction of Alzheimer's progression.
• Deploy the model in a user-friendly interface for clinical usage.

---

Acknowledgements

• We would like to thank the contributors of the dataset and the open-source libraries that made this project possible.
• We would also like to express our gratitude for the direction and guidelines from our project supervisor Dr. Muhammad Masroor Ali sir.

---

Contributors

• Md. Ishrak Ahsan
• Farhan Tahmidul Karim

Feel free to contribute, raise issues, or suggest enhancements! Reach out via ishrak26@gmail.com.