Influence of segmentation margin on machine learning-based high-dimensional quantitative CT texture analysis: a...

Area of Science:

• Radiology
• Medical Imaging
• Machine Learning in Oncology

Background:

• Quantitative computed tomography (CT) texture analysis (qCT-TA) is an emerging tool for characterizing renal clear cell carcinomas (RcCCs).
• The reproducibility and reliability of ML-based qCT-TA are crucial for clinical translation.
• Segmentation margin definition is a critical preprocessing step that can influence subsequent analysis.

Purpose of the Study:

• To investigate the impact of a 2mm segmentation margin variation on feature reproducibility, selection, and classification performance in ML-based qCT-TA of RcCCs.
• To compare contour-focused segmentation with margin shrinkage for qCT-TA of RcCCs.
• To assess the influence of segmentation on the classification of nuclear grade in RcCCs.

Main Methods:

• Retrospective analysis of 47 RcCC patient CT scans from a public database.
• Two segmentation methods: contour-focused and 2mm margin shrinkage.
• Extraction of texture features from original and filtered CT images, followed by correlation-based feature selection.
• Classification of nuclear grade (low vs. high) using k-nearest neighbors (KNN) ML classifier, with and without synthetic minority over-sampling technique (SMOTE).

Main Results:

• Margin shrinkage segmentation (2mm) yielded a higher proportion of reproducible features (93.2%) compared to contour-focused segmentation (86.2%) (p < 0.0001).
• Feature selection resulted in distinct feature subsets for each segmentation method, with only one common feature.
• ML models using contour-focused segmentation demonstrated superior classification performance (AUC range, 0.865-0.984) compared to margin shrinkage (AUC range, 0.745-0.887) (p < 0.05).

Conclusions:

• Each step of ML-based high-dimensional qCT-TA is sensitive to minor changes in segmentation margin.
• Contour-focused segmentation, despite yielding fewer highly reproducible features, resulted in better classification performance for distinguishing RcCC nuclear grade.
• The findings highlight the need to consider segmentation margin variability for reproducible ML-based qCT-TA in clinical practice.