Acute Myeloid Leukemia (AML) Subtype Classification Using Single Blood Cell Images

Project Overview

This project focuses on classifying subtypes of Acute Myeloid Leukemia (AML) using deep learning techniques. AML is a type of cancer that impacts the myeloid lineage of blood cells, often associated with specific genetic mutations. By leveraging single-cell blood smear images, the project aims to enhance diagnostic accuracy and support personalized treatment plans.

Table of Contents

1. Objectives
2. Dataset
3. Methodology
   • Single Cell Classification
   • Data Aggregation
   • AML Subtype Classification
4. Results
5. User Interface
6. Installation
7. Usage
8. References

Objectives

• Improve AML subtype classification accuracy through single-cell image analysis.
• Experiment with multiple deep learning architectures for sensitivity in detecting AML subtypes.
• Validate the model's diagnostic accuracy and real-world applicability to assist healthcare professionals.

Dataset

This project utilizes two main datasets:

1. Peripheral Blood Cell Images Dataset: Contains 17,092 high-resolution images of blood cells with various morphologies, annotated by pathologists.
2. AML and Control Group Dataset: Comprises 189 peripheral blood smears, divided by specific AML subtypes and a control group.

The data is preprocessed, ensuring class balance, and is used for training and testing deep learning models.

Methodology

Single Cell Classification

• A CNN model is trained on single-cell images to classify cell types such as neutrophils, lymphocytes, and platelets.
• Invalid image formats were handled with a custom data generator, improving data consistency for model training.

Data Aggregation

• The CNN model classifies each cell type within patient folders, creating a data frame of cell counts per patient.
• SMOTE is applied to balance classes and improve predictive performance.

AML Subtype Classification

• An ensemble approach (XGBoost, CatBoost, Random Forest, and Neural Network) is employed to predict AML subtypes.
• A binary classifier first identifies AML presence, and, if detected, the model further classifies the subtype.

Results

• The CNN achieved 94% accuracy on unseen data, indicating robust performance.
• Performance metrics across AML subtypes and control groups demonstrate high precision, recall, and F1-scores.
• Correlations of the single blood cell types with the presence/absence of AML suggest that basophils may indicate control, while lymphocytes and monocytes correlate with AML subtypes.

User Interface

The project includes a user-friendly interface that enables users to upload patient images for subtype classification, displaying cell type counts and AML predictions for diagnostic insight.