Lymph-vascular space invasion prediction in cervical cancer: Exploring radiomics and deep learning multilevel features of tumor and peritumor tissue on multiparametric MRI. (April 2020)

• Record Type: Journal Article
• Authors: Hua, Wenqing; Xiao, Taohui; Jiang, Xiran; Liu, Zaiyi; Wang, Meiyun; Zheng, Hairong; Wang, Shanshan
• Is Part Of: Biomedical signal processing and control. Volume 58(2020)
• Publication Date: 2020-04
• Subjects: Cervical cancer -- Radiomics -- Deep learning -- Lymph-vascular space invasion -- Magnetic resonance imaging; 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.2020.101869
• Languages: English
• ISSNs: 1746-8094

Abstract

Highlights: A radiomics and deep learning fusion model using multiparametric MRI was built for LVSI prediction in early-stage cervical cancer. Different from previous studies with focus on tumor region, the proposed method takes both tumor tissues and peri-tumor tissues with different radial dilation distances outside tumor in consideration for the prediction. We demonstrated that the peritumoral tissue contains remarkable information about the development process of LVSI in early-stage cervical cancer. Abstract: Preoperative determination of the presence of LVSI plays an important role in guiding surgical planning. In this paper, multiparametric magnetic resonance imaging (MRI)-based radiomics and deep feature learning strategy was applied to both tumor and peritumor tissues for preoperative prediction of LVSI in early-stage cervical cancer. 111 training cohort patients (44 LVSI-positive and 67 LVSI-negative) and 56 validation cohort patients (23 LVSI-positive and 33 LVSI-negative) with T1CE and T2WI modalities were enrolled. Radiomics features were extracted from tumor tissues, and peri-tumor tissues with different radial dilation distances outside tumor. The VGG-19 was used to extract high-level deep features. Support Vector Machine (SVM) models were constructed based on the radiomic and deep features extracted from multiparametric MRI. Models performance was evaluated on the validation cohort. Features extracted from tumor tissue with 8 mm and 4 mm radial dilation distances outside tumor show best discriminative performance for T1 CE and T2 WI respectively. For the final model construction, five radiomics features and three deep learning features were selected. The final model showed the best prediction results, with an AUC of 0.842 (95% confidence interval [CI], 0.772–0.913) in the training cohort and 0.775 (95% CI, 0.637–0.912) in the validation cohort. The sensitivity and specificity were 0.773 and 0.776 in the training cohort and 0.739 and 0.667 in the validation cohort. Taking into consideration of the features of peritumor tissues can contribute to improving LVSI prediction performance. The radiomics and deep learning fusion strategy shows the potential in prediction of LVSI in early-stage cervical cancer.