Multi-parametric MRI based radiomics with tumor subregion partitioning for differentiating benign and malignant soft-tissue tumors. (May 2021)

• Record Type: Journal Article
• Authors: Shang, Shengjie; Sun, Jing; Yue, Zhibin; Wang, Yingni; Wang, Xiaoyu; Luo, Yahong; Zhao, Dazhe; Yu, Tao; Jiang, Xiran
• Is Part Of: Biomedical signal processing and control. Volume 67(2021)
• Publication Date: 2021-05
• Subjects: STS soft-tissue tumor -- MRI magnetic resonance imaging -- CAD computer-aided diagnostic -- CNN Convolutional neural network -- ROI region of interest -- CE contrast-enhanced -- T1WI T1-weighted imaging -- KNN k-nearest neighbors -- ROC receiver operating characteristic -- AUC area under the ROC curves -- ACC accuracy -- SPE specificity -- SEN sensitivity; Soft-Tissue tumor -- MRI -- Radiomics -- Deep learning; Signal processing -- Periodicals; Biomedical engineering -- Periodicals; Signal Processing, Computer-Assisted -- Periodicals; Image Processing, Computer-Assisted -- Periodicals; Biomedical Engineering -- Periodicals; 610.28
• DOI: 10.1016/j.bspc.2021.102522
• Languages: English
• ISSNs: 1746-8094

Abstract

Highlights: Subregional radiomics was demonstrated to be effective on capturing the intratumoral heterogeneity in soft-tissue tumors. Handcrafted and deep learning-based features can provide complementary information to improve the diagnosis of STSs. A subregional radiomics nomogram was demonstrated to be a potential tool in aiding clinical diagnosis of STSs. Abstract: Objective: This study aims to explore MRI-based tumoral and intratumoral radiomics approaches on distinguishing malignant from benign soft-tissue tumors using handcrafted and deep learning-based features. Methods: A set of 82 patients underwent contrast-enhanced (CE) T1 and T1-weighted imaging (T1WI) MRI scans were enrolled between Jan. 2017 and Sep. 2019. The whole tumor regions were segmented by an unsupervised k -means algorithm. Radiomics handcrafted and deep learning-based features were extracted and selected from the whole tumor regions and intratumor subregions, and used to develop radiomics models based on a k -nearest neighbors (KNN) classifier. A radiomics nomogram was developed for potential clinical uses. The receiver operating characteristic (ROC) analysis was used to evaluate the discriminative performance of each model. Calibration and decision curve analysis were applied to evaluate the nomogram. Results: Our findings revealed that the active subregion in the CE-T1 and the whole tumor region in the T1WI MRI are the most discriminative regions. A fusion radiomics nomogram was established and achieved the best diagnostic performance with the area under the ROC curve (AUC) of 0.941 (SEN = 0.789, SPE = 1.000) in the training cohort and 0.922 (SEN = 0.667, SPE = 0.921) in the validation cohort. Conclusions: The proposed tumoral and intratumoral radiomics were potentially clinical valuable and could improve the application of computer-aided diagnosis (CAD) in soft-tissue tumor diagnosis.