ahmedhassan187/Image-Classification-of-Stroke-Blood-Clot-Origin

Image Classification of Stroke Blood Clot Origin Using Whole Slide Digital Pathology Images

Image Classification of Stroke Blood Clot Origin

Table of contents:

• Overview
• Data
• Model Stucture
• Results
• Project Structure
• Project Run
• Team

Overview

Problem related to STRIP-AI competition, to classify the blood clot origins in ischemic stroke. Using whole slide digital pathology images to differentiate between the two major acute ischemic stroke (AIS) etiology subtypes: cardiac and large artery atherosclerosis (CE - LAA).

Data

The data comprises 754 high-resolution whole-slide digital pathology images in TIF format, categorizing stroke origin as either CE or LAA, but have size variations and the distribution of the classes is not balanced that need careful handling.

To tackle the challenges of limited, imbalanced, and variable-sized stroke clot images, we leveraged data augmentation.

effectively balancing the dataset and boosting training data volume( 4313 final images).

Model Structure

Ensemble learning between ResNet and SqueezeNet

1. ResNet50 Model:

1. SqueezeNet Model:

Test Pipeline for one test image:

Results

In This part, we will see some results of each model we tried.

1. Confusion Matrix

• Confusion Matrix for results of SqueezeNet Model:

• Confusion Matrix for results of ResNet50 Model:

• Confusion Matrix for results of Ensemble Model:

2. ROC Curve

• ROC Curve for results of SqueezeNet Model with AUC = 0.65 :

• ROC Curve for results of ResNet50 Model with AUC = 0.69 :

• ROC Curve for results of Ensemble Model with AUC = 0.75 :

3. Other metrices

Metric	Score
Precision	0.833
Recall	0.714
F1 Score	0.769
Accuracy	0.700
Specificity	0.666

Project Structure

main
├─ Config
├─ data: dir containing 2 data samples
│  ├─  raw : original data
│  ├─  processed: Images Tiles
│  └─  processed-after-removing-low-score-imgs: Tiles after filtering 
├─ models: containing models checkpoints
├─ notebooks
│  ├─  Data Exploring.ipynb
│  ├─  Data Preprocessing Approach 1.ipynb
│  ├─  Data Preprocessing Approach 2.ipynb
│  ├─  EfficientNet_training_kaggle.ipynb
│  ├─  RestNet5_training_kaggle.ipynb
│  ├─  SqueezeNet_training_kaggle.ipynb
│  └─  test-evaluation.ipynb
├─ reports
│  ├─  Paper.pdf
|  └─   figures
├─ src
│  ├─  preprocessing
│  │   ├─  preprocessing_approach_1.py
│  │   └─  preprocessing_approach_2.py
|  ├─  dataloaders
│  │   ├─  Dataloader.py
│  │   └─  Dataloader_with_aug.py
│  └─  models
│      ├─  EfficientNet.py
│      ├─  ResNet50.py
│      ├─  squeezeNet.py
│      └─  helpers
│          └─  FireModule.py
└─ README.md

Project Run

Currently you can run and test project through these notebooks on Kaggle using competition data :

main
└─  notebooks
   ├─  Data Preprocessing Approach 2.ipynb
   ├─  RestNet5_training_kaggle.ipynb 
   ├─  SqueezeNet_training_kaggle.ipynb
   └─  test-evaluation.ipynb

Also, You can find models checkpoints on

• SqueezeNet
• ResNet50

Team

First Semester - Artificial Neural Networks in Medicine (SBE4025) class project created by:

Team Members' Names	Code
Ahmed Hassan	9202076
Habiba Fathallah	9202458
Rawan Mohamed	9202559
Romaisaa Saad	9202564

Submitted to:

• Dr. Inas Yassine & Eng. Merna Biabers All rights reserved © 2024 to Team 3 - Systems & Biomedical Engineering, Cairo University (Class 2024)