[Paper Review] Early Diagnosis of Acute Lymphoblastic Leukemia Using YOLOv8 and YOLOv11 Deep Learning Models

Alaa Awad, Salah A. Aly|arXiv (Cornell University)|Oct 14, 2024

Digital Imaging for Blood Diseases4 citations

TL;DR

This study proposes a novel early diagnosis framework for Acute Lymphoblastic Leukemia (ALL) using YOLOv8 and YOLOv11 deep learning models trained on merged multi-cell blood smear datasets. By leveraging transfer learning, data augmentation, and segmentation, the models achieve 98.8% testing accuracy in distinguishing malignant from benign white blood cells, including accurate detection of hematogones often misclassified as ALL.

ABSTRACT

Leukemia, a severe form of blood cancer, claims thousands of lives each year. This study focuses on the detection of Acute Lymphoblastic Leukemia (ALL) using advanced image processing and deep learning techniques. By leveraging recent advancements in artificial intelligence, the research evaluates the reliability of these methods in practical, real-world scenarios. Specifically, it examines the performance of state-of-the-art YOLO models, including YOLOv8 and YOLOv11, to distinguish between malignant and benign white blood cells and accurately identify different stages of ALL, including early stages. Moreover, the models demonstrate the ability to detect hematogones, which are frequently misclassified as ALL. With accuracy rates reaching 98.8%, this study highlights the potential of these algorithms to provide robust and precise leukemia detection across diverse datasets and conditions.

Motivation & Objective

• To develop a deep learning-based system for early and accurate detection of Acute Lymphoblastic Leukemia (ALL) in blood smear images.
• To address the gap in existing models by training on multi-cell images instead of single-cell datasets, reflecting real-world clinical conditions.
• To improve generalization by integrating two distinct datasets: the ALL image dataset and ALL-IDB1.
• To accurately classify hematogones as benign, reducing false positives commonly seen in prior models.
• To evaluate and compare the performance of YOLOv11 and YOLOv8 in detecting ALL across different stages and cell types.

Proposed method

• The study employs YOLOv11 and YOLOv8 small (s) versions as the primary object detection models for classifying white blood cells.
• Data preprocessing includes image segmentation, grayscaling, and masking to isolate individual cells from multi-cell blood smear images.
• Transfer learning and fine-tuning are applied using pre-trained YOLO architectures on the combined dataset of ALL-IDB1 and the ALL image dataset.
• The dataset is restructured into two classes: 'Normal' (merged benign from ALL dataset and normal from ALL-IDB1) and 'Cancer' (merged early, pre, and pro-phase from ALL dataset and cancerous from ALL-IDB1).
• Performance is evaluated using standard metrics: Accuracy = (TP + TN) / (TP + TN + FP + FN), Precision = TP / (TP + FP), Recall = TP / (TP + FN), F1-Score = 2 × (Precision × Recall) / (Precision + Recall), and Specificity = TN / (TN + FP).
• Training is conducted for 100 epochs with early stopping and learning rate scheduling to optimize convergence and generalization.

Experimental results

Research questions

• RQ1 Can YOLOv11 and YOLOv8 models achieve high accuracy in detecting early-stage Acute Lymphoblastic Leukemia from multi-cell blood smear images?
• RQ2 How does combining the ALL image dataset and ALL-IDB1 dataset improve model generalization across diverse sample types?
• RQ3 To what extent can the models correctly identify hematogones as non-malignant, reducing false positives in ALL diagnosis?
• RQ4 How do the performance metrics (accuracy, precision, recall, F1-score, specificity) of YOLOv11 compare to those of YOLOv8 in ALL detection?
• RQ5 What is the impact of data augmentation and transfer learning on model convergence and classification performance?

Key findings

• YOLOv11s achieved a testing accuracy of 98.8%, outperforming YOLOv8s, which reached 98.4% accuracy on the same dataset.
• The YOLOv11 model demonstrated stable training and validation loss curves with minimal fluctuations, indicating robust convergence.
• The confusion matrix revealed that only 0.06% of benign white blood cells (including hematogones) were misclassified as cancerous, highlighting high specificity.
• The integration of two diverse datasets (ALL image dataset and ALL-IDB1) significantly enhanced model generalization across different morphological patterns of blast cells.
• YOLOv11 achieved higher accuracy than previous state-of-the-art models such as VCaps-Net (98.64%) and custom CNNs (97.85%), establishing a new benchmark in ALL detection.
• The model successfully distinguished hematogones from malignant blasts, addressing a persistent challenge in prior ALL detection systems.