A Comparative Study of Texture Analysis Methods on the Classification Problem of DPLDs in HRCT scans.

Area of Science:

• Radiology and Medical Imaging
• Computer Vision
• Biomedical Engineering

Background:

• Texture classification is crucial for medical image analysis, particularly in identifying diffuse parenchymal lung diseases (DPLDs) from computed tomography (CT) scans.
• Challenges include similar patterns across instances and the need for invariance to rotation, scale, and lighting variations.
• Existing texture feature extraction methods require robust performance analysis for medical applications.

Purpose of the Study:

• To analyze and compare three classes of texture feature extraction methods for DPLDs in CT scans.
• To evaluate the performance of descriptive statistics, spectral analysis, and layered neural networks.
• To identify the most effective methods for accurate texture classification in this medical context.

Main Methods:

• Extracted texture features using descriptive statistics, spectral analysis, and layered neural networks (e.g., T-CNN, PCANet).
• Applied these feature extraction techniques to a dataset of CT scans featuring DPLDs.
• Conducted a performance analysis comparing the effectiveness of each class of methods.

Main Results:

• Performance analysis revealed significant differences in the efficacy of the tested texture feature extraction classes.
• T-CNN, PCANet, and Haralick Attributes demonstrated superior performance in texture characterization for DPLDs.
• Other methods within descriptive statistics and spectral analysis showed varied but generally lower performance.

Conclusions:

• T-CNN, PCANet, and Haralick Attributes are highly effective for texture classification of DPLDs in CT scans.
• These advanced methods offer improved accuracy and robustness compared to traditional techniques.
• The findings support the use of these advanced texture analysis methods in medical image diagnostics.