🤖 machine learning

Early Detection of Acute Myeloid Leukemia (AML) Using YOLOv12 Deep Learning Model

This study proposes an early detection framework for Acute Myeloid Leukemia (AML) using the YOLOv12 deep learning model combined with cell-based segmentation and Otsu thresholding preprocessing, achieving a validation and test accuracy of 99.3%.

Original authors: Enas E. Ahmed, Salah A. Aly, Mayar Moner

Published 2026-04-20

📖 4 min read☕ Coffee break read

CC0 1.0

Original authors: Enas E. Ahmed, Salah A. Aly, Mayar Moner

Original paper dedicated to the public domain under CC0 1.0 (http://creativecommons.org/publicdomain/zero/1.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine your body is a bustling city, and your blood is the highway system carrying millions of delivery trucks (blood cells). Usually, these trucks are well-organized: some carry oxygen, some fight infections, and some clean up waste.

But in Acute Myeloid Leukemia (AML), a dangerous glitch happens. The city's factory starts churning out "defective, immature trucks" (called myeloblasts) that clog the highways. These bad trucks look so much like the good ones that even expert human inspectors (doctors) can sometimes get confused when trying to sort them out under a microscope. If they miss the bad ones, the city (your body) can get overwhelmed very quickly.

This paper is about building a super-smart, automated robot inspector that never gets tired and can spot these bad trucks instantly.

Here is how the researchers built this robot, explained simply:

1. The Problem: Too Many Look-Alikes

The researchers had a huge photo album of blood cells. They needed to teach a computer to sort them into five different categories:

Myeloblasts: The dangerous, immature bad guys (the target).
Neutrophils, Monocytes, Basophils, Erythroblasts: The good guys (or at least, the normal ones).

The tricky part? These cells are tiny and look incredibly similar. It's like trying to tell the difference between five different shades of blue paint just by looking at a drop of each.

2. The Solution: A New Super-Eye (YOLOv12)

Instead of using old, slow cameras, the team used the latest and greatest AI model called YOLOv12.

The Analogy: Imagine an old security guard who has to stop every car, walk around it, check the tires, then the engine, then the driver before saying, "Okay, you can pass." That's slow.
YOLOv12 is like a super-fast drone flying over the highway. It looks at the entire road in a single glance and instantly shouts, "That's a bad truck! That's a good truck!" It does this so fast it can catch thousands of cars per second.

3. The Secret Sauce: Cleaning the Glasses

Before the robot could look at the photos, the researchers realized the photos were a bit "muddy." To help the robot see better, they tried two different ways to clean up the images (like putting on different pairs of glasses):

Glasses A (Hue Channel): This method filters the image to focus only on specific colors, like turning a color photo into a black-and-white sketch that highlights the shape.
Glasses B (Otsu Thresholding): This is like a high-contrast filter. It turns the image into a stark black-and-white silhouette, making the edges of the cells pop out sharply against the background.

They also tried two different ways of looking:

Looking at the whole cell (the entire truck).
Looking only at the nucleus (just the driver inside the truck).

4. The Big Race

The team ran a competition. They fed the robot photos processed by these different "glasses" and "viewpoints" to see which combination worked best.

The Result: The robot performed well in all scenarios, but the champion was the robot wearing the Otsu Thresholding glasses while looking at the whole cell.
The Score: This combination got a 99.3% accuracy.

What does 99.3% mean?
Imagine the robot inspecting 1,000 blood cells. It would correctly identify 993 of them. It only makes a mistake on 7. That is incredibly reliable for a life-or-death diagnosis.

5. Why This Matters

Speed: Because the AI is so fast, doctors could potentially screen thousands of patients in the time it takes to drink a cup of coffee.
Consistency: Humans get tired and their eyes get blurry. The robot doesn't. It sees the same thing every time.
Early Detection: By catching these "bad trucks" early, doctors can start treatment before the city (the patient's body) gets completely clogged up.

The Bottom Line

This paper is about teaching a super-fast AI robot to act as a tireless blood-cell detective. By using a new, advanced version of the YOLO model and cleaning up the images with a specific math trick (Otsu thresholding), they created a system that is almost perfect at spotting leukemia early. It's a major step toward making cancer diagnosis faster, cheaper, and more accurate for everyone.

1. Problem Statement

Acute Myeloid Leukemia (AML) is a life-threatening blood cancer characterized by the rapid proliferation of immature myeloid precursors (myeloblasts). Accurate diagnosis is critical but challenging due to:

Visual Similarity: Differentiating between various blood cell types (e.g., myeloblasts vs. segmented neutrophils) is difficult due to subtle morphological differences.
Data Limitations: Existing research suffers from a lack of well-annotated histopathological datasets and issues with class imbalance (specifically for rare immature cells like myeloblasts).
Methodological Flaws: Previous studies often lack independent test sets for unbiased evaluation or rely on models that struggle with generalization and reproducibility.

2. Methodology

The study proposes a comprehensive deep learning workflow consisting of three main phases: Data Preprocessing/Segmentation, Model Architecture, and Training/Evaluation.

A. Dataset

Source: A balanced, enhanced dataset derived from Kaggle, Mendeley Data, and The Cancer Imaging Archive (TCIA).
Content: 5,000 high-quality images (1024×1024 pixels, 1000x magnification, Wright-Giemsa stained) covering five critical cell types:
1. Myeloblasts (Hallmark of AML)
2. Segmented Neutrophils
3. Basophils
4. Monocytes
5. Erythroblasts
Split: 70% Training, 15% Validation, 15% Testing.

B. Preprocessing and Segmentation Strategies

To isolate relevant features, the authors applied two distinct segmentation techniques to generate four distinct datasets:

Hue Channel Filtering: Isolates cells based on color properties.
Otsu Thresholding: A statistical method to separate foreground (cells/nuclei) from the background.

Scope: Both techniques were applied to whole cells and nuclei separately, resulting in four input variations for the models.

C. Model Architecture: YOLOv12

The core innovation is the utilization of YOLOv12 (specifically the YOLOv12-N variant), the latest version of the "You Only Look Once" object detection framework.

Key Feature: Area Attention Mechanism. Unlike traditional attention mechanisms that are computationally expensive ( $O(n^2)$ ), YOLOv12 divides the feature map into horizontal or vertical segments (default $l=4$ ). This reduces complexity to $O(\frac{1}{2}n^2)$ , allowing the model to focus on critical image regions while maintaining real-time speed.
Comparison: The study notes YOLOv12 outperforms YOLOv10 and YOLOv11 in Mean Average Precision (mAP) while maintaining low latency (1.64ms on NVIDIA T4).
Baseline: The study also mentions the use of ResNet50 and Inception-ResNet50 v2 for comparison, though YOLOv12 yielded the superior results.

3. Key Contributions

Novel Architecture Application: First known application of YOLOv12 for AML cell classification, leveraging its new Area Attention mechanism for better feature extraction.
Robust Segmentation Pipeline: A systematic evaluation of two segmentation methods (Hue vs. Otsu) on two object levels (Cell vs. Nucleus) to determine the optimal preprocessing strategy.
Balanced Dataset Utilization: Addressing the issue of class imbalance by utilizing a curated, balanced dataset of 5,000 images across five distinct cell types.
Rigorous Evaluation: Strict separation of training, validation, and testing sets to ensure unbiased performance metrics and prevent data leakage.

4. Experimental Results

The models were evaluated using Accuracy, Sensitivity, Specificity, Precision, and F1-Score.

Best Performing Configuration: YOLOv12 with Otsu Thresholding on Cell-based segmentation.
- Validation Accuracy: 99.3%
- Test Accuracy: 99.3%
Other Configurations:
- Cell-Hue (YOLO): 98.8% (Val/Test)
- Nucleus-Hue (YOLO): 98.8% (Val/Test)
- Nucleus-Otsu (YOLO): 98.8% (Val/Test)
Observations:
- The Otsu thresholding method on whole cells provided the most distinct features for the model, leading to the highest accuracy.
- Training curves showed stable convergence, though minor fluctuations were observed in validation loss, which did not negatively impact final generalization.
- Confusion matrices confirmed high precision and reliable differentiation across all five cell classes.

5. Significance and Conclusion

This research demonstrates that integrating state-of-the-art object detection models (YOLOv12) with optimized image preprocessing (Otsu thresholding) can significantly enhance the early detection of AML.

Clinical Impact: The 99.3% accuracy suggests the system is highly reliable for assisting pathologists in distinguishing between malignant myeloblasts and normal blood cells, potentially reducing diagnostic errors and time.
Technical Advancement: The study validates the efficacy of the Area Attention mechanism in medical imaging, proving it can handle complex morphological features without sacrificing speed or accuracy.
Future Outlook: The authors suggest this pipeline serves as a scalable foundation for automated diagnostic tools, with potential for extension to other leukemia subtypes or integration into clinical workflows.