Introduction
The COVID-19 pandemic placed an unprecedented diagnostic burden on healthcare systems worldwide. While RT-PCR remains the primary diagnostic tool, Computed Tomography (CT) scans are vital for assessing infection severity and localization. However, manual evaluation of numerous 3D CT slices is time-intensive and susceptible to inter-observer variability. This paper proposes an automated image-processing pipeline that utilizes Harmony Search Optimization (HSO) and Otsu thresholding to quantify pneumonia infection levels, facilitating rapid patient triaging.
Methodology
The proposed system employs a four-stage computational pipeline to isolate and measure infected lung tissue.
- Preprocessing: A manual threshold filter is initially applied to remove non-pulmonary artifacts, such as bone and surrounding soft tissue, from the CT slices.
- Image Enhancement: The core of the system involves multi-level thresholding. The researchers utilize Otsu’s method, optimized via the Harmony Search Optimization (HSO) algorithm. HSO, a metaheuristic inspired by musical improvisation, efficiently identifies optimal threshold levels to distinguish between healthy tissue, ground-glass opacities (GGOs), and background noise.
- Segmentation: A Watershed algorithm is then used to define the boundaries of the infected regions based on the enhanced image topography.
- Severity Computation: The system calculates a "Severity-Level-Measure" (SLM) or Percentage Infection Rate by comparing the pixel density of the extracted infected sections to the total lung area: $$\text{Infection Rate} = \left( \frac{\text{Pixel density of infected section}}{\text{Pixel density of total lung segment}} \right) \times 100$$
Experimental Setup and Datasets
The study utilized clinical-grade CT images sourced from the Radiopaedia open-access database. The dataset comprised 110 2D slices across 11 patients, including both coronal (90 slices) and axial (20 slices) views.
Results and Discussion
The system demonstrated high efficiency and clinical robustness. Key findings include:
- Processing Speed: The mean processing time was approximately 38 seconds for a set of 10 slices per patient, significantly faster than manual assessment.
- Orientation Independence: The algorithm yielded consistent results across different imaging planes. For instance, a 75-year-old male patient showed nearly identical infection rates in coronal (38.52%) and axial (39.06%) views.
- Severity Mapping: The tool successfully quantified varying degrees of infection, ranging from mild cases (e.g., 9.95% in Case 39) to severe cases (e.g., 32.02% in Case 15).
Conclusion
The integration of HSO with Otsu thresholding provides a fast, objective metric for COVID-19 severity. By quantifying the percentage of lung involvement, this system assists radiologists in prioritizing high-risk patients. Future work will focus on expanding the dataset, automating the classification of severity levels (Mild, Moderate, Severe), and distinguishing COVID-19 from other pulmonary infections.
Summary updated