Robustness-Driven Feature Selection in Classification of Fibrotic Interstitial Lung Disease Patterns in Computed...

Area of Science:

• Medical Imaging
• Artificial Intelligence in Medicine
• Radiology

Background:

• Computer-aided diagnosis (CADx) systems for computed tomography (CT) face adoption barriers due to classifier fragility.
• Variations in CT technical factors (e.g., slice thickness, reconstruction kernel, tube current) significantly impact classifier robustness.

Purpose of the Study:

• To introduce and evaluate a novel Robustness-Driven Feature Selection (RDFS) algorithm.
• To enhance the robustness of CT classification models against technical variations.

Main Methods:

• Developed and applied the RDFS algorithm to select features resilient to CT technical variations.
• Evaluated RDFS using 3D texture features for classifying fibrotic interstitial lung disease in 99 adult subjects.
• Compared two support vector machine classifiers: one with RDFS and one without, using multi-reconstruction and testing datasets.

Main Results:

• The RDFS-enhanced classifier demonstrated superior robustness (Cohen's kappa 0.899-0.989) compared to the non-RDFS classifier (kappa 0.827-0.968).
• Classifier performance was comparable between the two methods on the testing dataset (EGM 0.778 with RDFS vs. 0.785 without RDFS).
• RDFS effectively improved robustness against variations in slice thickness, reconstruction kernel, and tube current without performance degradation.

Conclusions:

• RDFS significantly enhances classifier robustness for CT-based computer-aided diagnosis.
• The algorithm is effective in maintaining classifier performance while mitigating the impact of technical scan variations.
• RDFS has implications for multicenter clinical trials requiring reproducible quantitative CT image analysis across diverse settings.