Disease Localization and Severity Assessment in Chest X-Ray Images using Multi-Stage Superpixels Classification

Area of Science:

• Medical Imaging Analysis
• Artificial Intelligence in Healthcare
• Radiomics and Computational Pathology

Background:

• Chest X-rays (CXRs) are crucial for managing chronic lung diseases like tuberculosis and pneumonia.
• Existing computer-aided diagnosis (CAD) systems primarily focus on classification, with limited capabilities in precise disease localization and severity assessment.
• Current deep learning methods for localization often produce imprecise results using saliency maps.

Purpose of the Study:

• To develop a robust framework for generating compact disease boundaries and infection maps from CXRs.
• To accurately grade the severity of lung infections using a multistage superpixel classification approach.
• To enhance the diagnostic capabilities of CAD systems for lung diseases.

Main Methods:

• Utilized Simple Linear Iterative Clustering (SLIC) to segment lung fields into superpixels.
• Extracted and combined radiomic texture and shape features for multistage classifier training.
• Generated infection maps, disease boundaries, and severity grades based on predicted superpixel labels.

Main Results:

• Achieved high accuracy (95.52%) and F-Measure (95.48%) in Stage-I classification on the calibration dataset.
• Demonstrated strong performance on the validation dataset with accuracy up to 93.41% and F-Measure of 95.32%.
• Obtained an average Jaccard Index of 0.82 for localization and a high correlation (r=0.9589) between predicted and radiologist severity scores.

Conclusions:

• The proposed framework shows promising performance for disease localization and severity grading in CXRs.
• The high Jaccard Index and correlation coefficient indicate the robustness and clinical potential of the method.
• Statistical validation confirms the significance of the results, supporting its utility in computer-aided diagnosis.