Classification of interstitial lung disease patterns using local DCT features and random forest

Area of Science:

• Medical Imaging
• Computer-Aided Diagnosis
• Pulmonary Medicine

Background:

• Interstitial lung diseases (ILDs) pose diagnostic challenges, necessitating advanced tools for accurate pattern identification.
• Computer-aided diagnosis (CAD) systems have been developed to aid physicians in diagnosing ILDs from high-resolution computed tomography (HRCT) images.
• Accurate classification of ILD patterns is crucial for disease quantification and management.

Purpose of the Study:

• To propose a novel scheme for classifying HRCT image patches exhibiting ILD abnormalities.
• To develop a foundational component for the quantitative analysis of diverse ILD patterns.
• To enhance the accuracy and efficiency of ILD diagnosis through advanced image analysis techniques.

Main Methods:

• Feature extraction using local spectral analysis with a Discrete Cosine Transform (DCT)-based filter bank.
• Characterization of image texture by computing q-quantiles to describe local frequency distributions.
• Integration of gray-level histogram values to form a comprehensive feature vector.
• Classification of image patches using a Random Forest (RF) classifier.

Main Results:

• The proposed approach demonstrated superior performance in classifying HRCT image patches with ILD abnormalities.
• Experimental results validated the efficiency of the developed feature extraction and classification scheme.
• The method achieved state-of-the-art results compared to existing approaches.

Conclusions:

• The developed scheme offers a robust and efficient method for the classification of ILD patterns in HRCT images.
• This approach serves as a key component for the quantification of ILD, potentially improving diagnostic outcomes.
• The findings highlight the potential of spectral analysis and machine learning in advancing pulmonary diagnostics.