Introduction
Manual leukocyte analysis is a significant bottleneck in medical diagnostics. It is labor-intensive, time-consuming, and prone to inter-observer variability. While automation via computer vision is a goal, blood smear images present challenges: they are often noisy, cells frequently overlap or are deformed, and contrast between cell boundaries and the background can be poor. Traditional tools like the Hough Transform are computationally expensive and sensitive to noise, while AI methods like Genetic Algorithms often struggle with local optima and complex hyperparameter tuning.
This paper introduces a novel approach by adapting the Plant Growth Simulation Algorithm (PGSA) to detect leukocytes, treating the task as a circle-detection problem. This bionic strategy, inspired by how plants grow toward light, provides a robust and efficient detection system.
Methodology
The proposed system follows a multi-stage pipeline designed to identify white blood cells as "quasi-circular" objects.
- Preprocessing: Histogram thresholding segments leukocytes from red blood cells and the background. Morphological edge detection (erosion and inversion) then creates a clean map of cell boundaries.
- The PGSA Framework: The algorithm mimics phototropism. The "light source" represents the global optimal solution (the actual cell). Growth points (candidate solutions) are assigned a "morphactin concentration" based on their fitness—how closely they resemble a cell. Higher concentration encourages further "branching" or searching from that point.
- Circle Fitting: Three random points on the edge map define a candidate circle. The PGSA evolves these candidates through successive iterations, measuring resemblance to actual leukocytes via a fitness function.
Experimental Setup
The study utilized a dataset of 80 microscopic blood smear images from Cellavision.com (360 x 363 pixels). The dataset contained 463 white blood cells, including both "bright" and "dark" leukocytes to test staining intensity variations. A human hematologist provided the ground truth for all cells.
Results and Discussion
The PGSA-based method demonstrates significant advantages in both accuracy and efficiency. By mimicking natural growth, the algorithm reduces the search space in each iteration, leading to lower computational runtime than exhaustive searches. It also functions with subpixel precision, identifying boundaries more accurately than pixel-grid-limited methods.
Compared against the Hough Transform (HT), Genetic Algorithms with Ant Colony Optimization (GA+ACO), and Fuzzy Cellular Neural Networks (FCCN), the proposed method achieved:
- Accuracy: A 98.28% True Positive Rate and a low 1.72% False Positive Rate.
- Noise Robustness: The system maintained high precision even under 10% Salt & Pepper noise, whereas the Hough Transform's performance dropped to ~24%.
- Visual Success: The algorithm successfully identified cells that were deformed, partially occluded, or embedded in cluttered backgrounds.
Conclusion
While highly effective—even with cells up to 68% hidden—the algorithm still faces challenges with leukocytes that are almost entirely obscured. Future work will focus on detecting these completely hidden cells and extending the PGSA framework to other circular medical features, such as ovarian or liver cells.
Summary updated