Influence of segmentation margin on machine learning-based high-dimensional quantitative CT texture analysis: a reproducibility study on renal clear cell carcinomas.

OBJECTIVE: To determine the possible influence of segmentation margin on each step (feature reproducibility, selection, and classification) of the machine learning (ML)-based high-dimensional quantitative computed tomography (CT) texture analysis (qCT-TA) of renal clear cell carcinomas (RcCCs).

MATERIALS AND METHODS: For this retrospective study, 47 patients with RcCC were included from a public database. Two segmentations were obtained by two radiologists for each tumour: (i) contour-focused and (ii) margin shrinkage of 2 mm. Texture features were extracted from original, filtered, and transformed CT images. Feature selection was done using a correlation-based algorithm. The ML classifier was k-nearest neighbours. Classifications were performed with and without using synthetic minority over-sampling technique. Reference standard was nuclear grade (low versus high). Intraclass correlation coefficient (ICC), Pearson's correlation coefficient, Wilcoxon signed-ranks test, and McNemar's test were used in the analysis.

RESULTS: The segmentation with margin shrinkage of 2 mm (772 of 828; 93.2%) yielded more texture features with excellent reproducibility (ICC ≥ 0.9) than contour-focused segmentation (714 of 828; 86.2%), p < 0.0001. The feature selection algorithms resulted in different feature subsets for two segmentation datasets with only one common feature. All ML-based models based on contour-focused segmentation (area under the curve [AUC] range, 0.865-0.984) performed better than those with margin shrinkage of 2 mm (AUC range, 0.745-0.887), p < 0.05.

CONCLUSIONS: Each step of the ML-based high-dimensional qCT-TA was susceptible to a slight change of 2 mm in segmentation margin. Despite yielding fewer features with excellent reproducibility, use of the contour-focused segmentation provided better classification performance for distinguishing nuclear grade.

KEY POINTS: • Each step of a machine learning (ML)-based high-dimensional quantitative computed tomography texture analysis (qCT-TA) is sensitive to even a slight change of 2 mm in segmentation margin. • Despite yielding fewer texture features with excellent reproducibility, performing the segmentation focusing on the outermost boundary of the tumours provides better classification performance in ML-based qCT-TA of renal clear cell carcinomas for distinguishing nuclear grade. • Findings of an ML-based high-dimensional qCT-TA may not be reproducible in clinical practice even using the same feature selection algorithm and ML classifier unless the possible influence of the segmentation margin is considered.