Repeatability of Radiomic Features Against Simulated Scanning Position Stochasticity Across Imaging Modalities and Cancer Subtypes: A Retrospective Multi-institutional Study on Head-and-Neck Cases

• Published in: Computational Mathematics Modeling in Cancer Analysis Vol. 13574; pp. 21 - 34
• Main Authors: Zhang, Jiang, Lam, Saikit, Teng, Xinzhi, Zhang, Yuanpeng, Ma, Zongrui, Lee, Francis, Au, Kwok-hung, Yip, Wai Yi, Chang, Tien Yee Amy, Chan, Wing Chi Lawrence, Lee, Victor, Wu, Q. Jackie, Cai, Jing
• Format: Book Chapter
• Language: English
• Published: Switzerland Springer 2022; Springer Nature Switzerland
• Series: Lecture Notes in Computer Science
• Subjects: Head and neck cancer; Radomics; Repeatability
• ISBN: 9783031172656; 3031172655
• ISSN: 0302-9743; 1611-3349
• DOI: 10.1007/978-3-031-17266-3_3

We attempted to investigate the Radiomic feature (RF) repeatability and its agreements across imaging modalities and head-and-neck cancer (HNC) subtypes via image perturbations. Contrast-enhanced computed tomography (CECT), CET1-weight, T2-weight magnetic resonance images of 231 nasopharyngeal carcinoma (NPC) patients, and CECT images of 399 oropharyngeal carcinoma (OPC) patients were retrospectively analyzed. Randomized translation and rotation were implemented to the images for mimicking scanning position stochasticity. 1288 RFs from unfiltered, Laplacian-of-Gaussian-filtered (LoG), and wavelet-filtered images were subsequently computed per perturbed image. The intra-class correlation coefficient (ICC) was calculated to assess RF repeatability. The mean absolute difference (MAD) of the ICC and the binarized repeatability consistency between image sets were adopted to evaluate its agreements across imaging modalities and HNC subtypes. Bias from feature collinearity was also investigated. All the shape RFs and the majority of RFs from unfiltered (≥ $$\ge $$ 83.5%) and LoG-filtered (≥ $$\ge $$ 93%) images showed high repeatability (ICC ≥ $$\ge $$ 0.9) in all studied datasets, whereas more than 50% of the wavelet-filtered RFs had low repeatability (ICC < 0.9). RF repeatability agreements between imaging modalities within the NPC cohort were outstanding (MAD < 0.05, consistency > 0.9) and slightly higher between the NPC and OPC cohort (MAD = 0.06, consistency = 0.89). Minimum bias from feature collinearity was observed. We urge caution when handling wavelet-filtered RFs and advise taking initiatives to exclude underperforming RFs during feature pre-selection for robust model construction.