An Unsupervised Correlation Learning-Based Clustering Model for Multiple Complex Lesions Evaluation.

Area of Science:

• Medical Imaging Analysis
• Machine Learning in Healthcare
• Computational Pathology

Background:

• Accurate evaluation of lesion morphology and quantity in computed tomography (CT) images is vital for disease diagnosis.
• Current machine learning methods often analyze lesion morphology and quantity separately, failing to capture the synergistic relationship crucial for complex cases with multiple lesions.

Purpose of the Study:

• To propose an unsupervised correlation learning-based clustering model for evaluating both morphology and quantity of multiple complex lesions in CT images.
• To address the limitations of existing methods by integrating morphological structure and quantitative distribution analysis without predefined logic.

Main Methods:

• Developed an unsupervised model utilizing clinical knowledge and lesion region in/out-degree to learn interdependencies and recognize domain-specific morphological features.
• Perceived quantity evaluation as a density-based clustering process, dynamically adjusting search based on morphological features and employing morphology-special parameter search strategies.
• Validated the model on kidney stone and kidney tumor datasets.

Main Results:

• Achieved 92.45% accuracy in morphology analysis for kidney stones and 95.33% for kidney tumors.
• Attained 79.25% accuracy in quantity analysis for kidney stones and 94.33% for kidney tumors.
• Outperformed AR-DBSCAN by +30.19% and DRL-DBSCAN by +6% in quantity analysis.

Conclusions:

• The proposed unsupervised correlation learning model effectively handles morphology and quantity estimation for multiple complex lesions in CT imaging.
• The model demonstrates superior performance compared to existing methods, offering a robust solution for intricate diagnostic scenarios.
• The integration of morphological and quantitative analysis provides a more comprehensive approach to lesion evaluation.