An attention-based deep learning for acute lymphoblastic leukemia classification

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Hematology

Background:

• Acute Lymphoblastic Leukemia (ALL) is a bone marrow malignancy characterized by overproduction of immature cells.
• ALL diagnoses in the U.S. approach 6500 annually, representing a significant portion of pediatric cancers.
• Computer-assisted diagnosis (CAD) systems are increasingly vital for hematologists to manage data and improve diagnostic accuracy.

Purpose of the Study:

• To introduce a novel Deep Dilated Residual Convolutional Neural Network (DDRNet) for automated blood cell classification.
• To enhance feature extraction and classification accuracy for early detection of Acute Lymphoblastic Leukemia (ALL).
• To address challenges in CAD systems, such as vanishing gradients and feature discrimination.

Main Methods:

• The study developed a DDRNet model incorporating Deep Residual Dilated Blocks (DRDB), Global and Local Feature Enhancement Blocks (GLFEB), and Channel and Spatial Attention Blocks (CSAB).
• The model utilizes Tanh and sigmoid activation functions for non-linearity and feature concentration.
• A Kaggle dataset of 16,249 blood cell images, categorized into four classes, was used for training (80%) and testing (20%).

Main Results:

• The DDRNet model achieved a high training accuracy of 99.86% and a testing accuracy of 91.98%.
• The model demonstrated a significant F1 score of 0.96, indicating robust classification performance.
• The integration of DRDB, GLFEB, and CSAB blocks enhanced feature discrimination with minimal computational complexity.

Conclusions:

• The DDRNet model offers superior performance in multi-class blood cell classification compared to existing methods.
• The proposed architecture effectively addresses challenges in feature extraction and classification for ALL detection.
• DDRNet shows promise as an advanced CAD tool for hematologists, improving diagnostic efficiency and accuracy.