Content-based image retrieval for Lung Nodule Classification Using Texture Features and Learned Distance Metric

Area of Science:

• Medical Imaging
• Computer-Aided Diagnosis
• Artificial Intelligence

Background:

• Accurate differentiation of benign and malignant lung nodules on CT scans is crucial for patient management.
• Content-based image retrieval (CBIR) offers a promising approach for analyzing lung nodule characteristics.

Purpose of the Study:

• To introduce a novel two-step CBIR scheme (TSCBIR) for enhanced computer-aided diagnosis of lung nodules.
• To develop and evaluate new similarity metrics for lung nodule comparison.

Main Methods:

• The proposed TSCBIR scheme utilizes two similarity metrics: semantic relevance and visual similarity.
• The first step involves retrieving K most similar regions of interest (ROIs) based on semantic relevance.
• The second step weights retrieved ROIs by visual similarity to predict malignancy likelihood.

Main Results:

• A dataset of 366 nodule ROIs from LIDC-IDRI CT scans was utilized for validation.
• Three texture feature groups were employed to represent nodule ROIs.
• The TSCBIR scheme demonstrated significant performance improvements compared to existing classifiers.

Conclusions:

• The proposed TSCBIR method effectively measures nodule similarity for improved lung nodule diagnosis.
• This approach shows potential for enhancing the accuracy of differentiating malignant from benign lung lesions on CT images.