Radiomics analysis of apparent diffusion coefficient in cervical cancer: A preliminary study on histological grade evaluation

• Published in: Journal of magnetic resonance imaging Vol. 49; no. 1; pp. 280 - 290
• Main Authors: Liu, Ying, Zhang, Yuwei, Cheng, Runfen, Liu, Shichang, Qu, Fangyuan, Yin, Xiaoyu, Wang, Qin, Xiao, Bohan, Ye, Zhaoxiang
• Format: Journal Article
• Language: English
• Published: United States Wiley Subscription Services, Inc 01.01.2019
• Subjects: Adult; Aged; Aged, 80 and over; apparent diffusion coefficient; Biopsy; Cancer; Carcinoma, Squamous Cell - diagnostic imaging; Cervical cancer; Diffusion; Diffusion coefficient; Diffusion Magnetic Resonance Imaging; diffusion‐weighted magnetic resonance imaging; Feature extraction; Female; Field strength; histopathological grade; Human papillomavirus; Humans; Image acquisition; Image Interpretation, Computer-Assisted - methods; Image processing; Image Processing, Computer-Assisted - methods; Image segmentation; Magnetic resonance imaging; Mathematical models; Medical imaging; Middle Aged; Neoplasm Grading; Parameters; Patients; Prospective Studies; Quality; Radiomics; Rank tests; Regression analysis; Reproducibility; Reproducibility of Results; Squamous cell carcinoma; Statistical analysis; Statistical tests; Three dimensional models; Tumors; Two dimensional models; Uterine Cervical Neoplasms - diagnostic imaging; Uterus; apparent diffusion coefficient; radiomics; histopathological grade; cervical cancer; diffusion-weighted magnetic resonance imaging
• DOI: 10.1002/jmri.26192
• ISSN: 1053-1807; 1522-2586

Background The role of apparent diffusion coefficient (ADC)‐based radiomics features in evaluating histopathological grade of cervical cancer is unresolved. Purpose To determine if there is a difference between radiomics features derived from center‐slice 2D versus whole‐tumor volumetric 3D for ADC measurements in patients with cervical cancer regarding tumor histopathological grade, and systematically assess the impact of the b value on radiomics analysis in ADC quantifications. Study Type Prospective. Subjects In all, 160 patients with histopathologically confirmed squamous cell carcinoma of uterine cervix. Field Strength/Sequence Conventional and diffusion‐weighted MR images (b values = 0, 800, 1000 s/mm2) were acquired on a 3.0T MR scanner. Assessment Regions of interest (ROIs) were drawn manually along the margin of tumor on each slice, and then the center slice of the tumor was selected with naked eyes in the course of whole‐tumor segmentation. A total of 624 radiomics features were derived from T2‐weighted images and ADC maps. We randomly selected 50 cases and did the reproducibility analysis. Statistical Tests Parameters were compared using Wilcoxon signed rank test, Bland–Altman analysis, t‐test, and least absolute shrinkage and selection operator (LASSO) regression with crossvalidation. Results In all, 95 radiomics features were insensitive to ROI variation among T2 images, ADC map of b800, and ADC map of b1000 (P > 0.0002). There was a significant statistical difference between the performances of 2D center‐slice and 3D whole‐tumor radiomics models in both ADC feature sets of b800 and b1000 (P < 0.0001, P < 0.0001). Compared with ADC features of b800 (0.3758 ± 0.0118), the model of b1000 ADC features appeared to be slightly lower in overall misclassification error (0.3642 ± 0.0162) (P = 0.0076). Data Conclusion Several radiomics features extracted from T2 images and ADC maps were highly reproducible. Whole‐tumor volumetric 3D radiomics analysis had a better performance than using the 2D center‐slice of tumor in stratifying the histological grade of cervical cancer. A b value of 1000 s/mm2 is suggested as the optimal parameter in pelvic DWI scans. Level of Evidence: 1 Technical Efficacy: Stage 1 J. Magn. Reson. Imaging 2019;49:280–290.