1713
Prediction of stage, differentiation and Ki-67 status of locally advanced cervical cancer by DCE-MRI texture analysis1Radiology, Central Hospital of Wuhan, Tongji Medical College, Huazhong University of Science and Technology, Wuhan, China, 2Central Hospital of Wuhan, Tongji Medical College, Huazhong University of Science and Technology, Wuhan, China, 3Philips Healthcare, Shanghai, China
Synopsis
This retrospective study explored the value of texture
analysis in predicting the stage, differentiation and Ki-67 status of pretreatment
advanced cervical cancer. Multi-class radiomics feature extraction was
performed on the maximum enhancement (ME) and maximum relative enhancement
(MRE) maps from DCE-MRI. A prediction model using a machine
learning-XGB classifier showed the mean
sensitivities of predicting FIGOⅡb-Ⅲa, poor differentiation and high Ki-67 status were 0.767,
0.963 and 0.967; specificities were 0.958, 0.361 and 0.694 , and AUCs were
0.910, 0.920 and 0.840 respectively. DCE-MRI textural parameters have potential
as non-invasive imaging biomarkers in predicting histopathology in advanced
cervical cancer.
Introduction
Cervical cancer is the fourth most common cause of death by cancer in the Western world and the second most common cause of death in the developing world1. Accurate staging and tumor grading has led to more efficient therapies with better outcomes and reduced side effects: while patients with low-grade and locally confined tumors (FIGO stages IB1 and IIA1) may be amenable to surgical resection alone, advanced and poorly differentiated tumors require combined neoadjuvant chemotherapy and/ or radio-chemotherapy. FDG-PET showed a varied FDG uptake in cervical cancer by histology and differentiation, which both well and poorly differentiated tumors also had a higher SUVmax2. Dynamic contrast-enhanced MRI (DCE-MRI) has been employed to evaluate the extent of tumor angiogenesis and tumor heterogeneity by analyzing patterns of enhancement3. Many studies have explored heterogeneous enhancement patterns in DCE-MR images within the entire tumor to build predictive models of tumor subtypes based on the quantitative evaluation of contrast enhancement4. Texture analysis is a mathematical statistical procedure to extract objective and quantitative parameters (texture features) from given images. Several studies showed that texture features derived from DWI or DCE-MRI potentially predicted histological tumor differentiation and cancer stage5,6. Here, we report a potential method, DCE- MRI maps texture analysis combined with clinical index for evaluating cervical cancer. In this retrospective study, the first order and three higher order features (GLRM, GLRLM and GLSZM) were used for analyzing whole tumor DCE-MRI maps to investigate the value of advanced stage, poor differentiation and higher Ki-67 status of pretreatment cervical cancer.Purpose
To determine the value of texture analysis of DCE-MRI maps in prediction of advanced stage, poor differentiation and high Ki-67 status of advanced cervical cancer.Methods
Thirty-nine patients were enrolled in this retrospective, institutional review board (IRB)-approved study. DCE-MRI was performed on 3.0T scanner (Ingenia, Phillip Healthcare, Best, the Netherlands) by a 3D T1W High Resolution Isotropic Volume Examination (THRIVE) sequence with 36 phases. Co-occurrence matrix -based texture features were extracted from each tumor on maximum enhancement (ME) and maximum relative enhancement (MRE) maps from DCE-MRI using in-home radiomics tool based on Mat-Lab software. Multivariate models were trained on the training cohort and their performance was evaluated on the 5-fold cross-validation cohort using the area under ROC curve (AUC), accuracy, specificity and sensitivity. P value<0.05 was considered statistically significant.Results
Mean age was 56.5±10.3 years. Histopathology revealed 9 adenocarcinoma and 30 squamous cell cancer; 7 well-differentiated, 21 moderately differentiated or moderately to poorly differentiated, and 11 poorly differentiated tumors; 7 FIGO Ⅰb, 18 FIGO Ⅱa, 8 FIGO Ⅱb, 6 FIGO Ⅲa, none of FIGO Ⅲb and Ⅳ. Lymph nodes were involved in 12 patients. On ME maps, two GLSZM, two GLRLM and one GLCM features correlated with the grades: LAHGE (r=0.37, P=0.02), GLN(GLSZM) (r=0.386, P=0.015), GLN(GLRLM) (r=0.325, P=0.044), RLN (r=0.444, P=0.005), correlation(GLCM) (r=0.467, P=0.003); On MRE maps, three GLSZM and two GLRLM features correlated with the grades: LAHGE (r=0.38, P=0.017), GLN (GLSZM) (r=0.354, P=0.027), and SZN (r=0.332, P=0.039), GLN(GLRLM) (r=0.392, P=0.014), RLN (r=0.361, P=0.024) respectively. Among all texture features derived from ME and MRE maps, correlation (GLCM) on ME showed a correlation with tumor differentiation. We used a machine learning-XGB classifier to build a prediction model using all the radiomics features. Twenty-nine features were chosen to build the model and results showed the mean sensitivities of predicting FIGOⅡb-Ⅲa, poor differentiation and high Ki-67 status were 0.767, 0.963 and 0.967 respectively; specificities were 0.958, 0.361 and 0.694 , and AUCs were 0.910, 0.920 and 0.840 respectively. The ROC curves were shown in Figure 2-4.Conclusion and discussion
Volumetric texture analysis on DCE-MRI maps can potentially help to predict tumor with poorly differentiation, higher stages and Ki-67 status for cervical cancer stage FIGO Ⅰb-Ⅲa. In this study, we used 1765 features inclusing first order and three kinds of texture features (GLRM, GLRLM and GLSZM) to analyze whole tumor DCE-MRI maps. Our results showed several texture futures were associated with differentiation, FIGO stages respectively. It remain a controversial point that not be found on an ADC map texture study5.Heterogeneous and varied enhanced patterns in intratumoral regions correlated with clinical and histologic features, might be the dominated explanations for the difference of texture features among tumor stages, grades and differentiations respectively. The cross-validation performance of the trained model in the present study showed a relatively high accuracy, indicating that the large number of support vectors may simply reflect the considerable variation in tumor characteristics among the patients. However, the statistical power was limited due to the relatively small number of samples. Further research will be necessary to verify our preliminary findings in a larger cohort.Acknowledgements
No acknowledgement found.References
1. Lea JS, Lin KY. Cervical cancer. Obstetrics and gynecology clinics of North America. 2012;39(2):233-253.
2. Kidd EA, Spencer CR, Huettner PC, Siegel BA, Dehdashti F, Rader JS, et al. Cervical cancer histology and tumor differentiation affect 18F-fluorodeoxyglucose uptake. Cancer. 2009;115(15):3548-3554.
3.Yuh WT, Mayr NA, Jarjoura D, Wu D, Grecula JC, Lo SS, et al. Predicting control of primary tumor and survival by DCE MRI during early therapy in cervical cancer. Investigative radiology. 2009;44(6):343-50. 4.M. A. Zahra, K. G. Hollingsworth, E. Sala, D. J.Lomas, and L. T. Tan, “Dynamic contrast-enhanced MRI as a predictor of tumour response to radiotherapy,” Lancet Oncology,2007;8(1): 63-74.
5.Guan Y, Li W, Jiang Z, Zhang B, Chen Y, Huang X, et al. Value of whole-lesion apparent diffusion coefficient (ADC) first-order statistics and texture features in clinical staging of cervical cancers. Clinical radiology. 2017;72(11):951-958.
6.Torheim T, Malinen E, Kvaal K, Lyng H, Indahl UG, Andersen EK, et al. Classification of dynamic contrast enhanced MR images of cervical cancers using texture analysis and support vector machines. IEEE transactions on medical imaging. 2014;33(8):1648-1656.
Figures
