Preclinical And Clinical Reasearch On MR-based Habitat Imaging Radiomics As Pretreatment Predictors Of Intratumoral Heterogeneity Of Treatment Responses And Prognosis To Concurrent Chemo-Radiotherapy In Locally Advanced Cervical Cancer

Part I Exploring the Value of Magnetic Resonance Habitat Imaging and Radiomics Prediction of early Treatment Response and correlation with Prognosis of Concurrent Chemoradiotherapy for Locally Advanced Cervical CancerObjective: To construct a machine learning model based on Contrast Enhanced-MR Habitat Imaging and Radiomics to predict the early treatment response and long-term prognostic relevance of patients treated with concurrent chemoradiotherapy in patients with Locally Advanced Cervical Cancer.Methods: A total of 268 patients with locally advanced cervical squamous cancer treated with radical concurrent radiotherapy between May 2013 and March 2021 were finally included in this retrospective study.All patients enrolled underwent CE-MR within 2 weeks prior to treatment and acquired both T1 WI and T2 WI and contrast-enhanced T1-WI(early and late phase of arterial enhancement).Early treatment response was assessed according to RECIST criteria(CR and PR were defined as early efficacy response group;SD and PD were defined as early efficacy non-response group);long-term prognosis was assessed using progression-free survival(PFS)and overall survival(OS).Semi-automatic manual segmentation of ROI/VOI of the primary tumor was performed in MR-T2-weighted images.Firstly,we performed the feature downscaling by Principal Component Analysis(PCA)and feature filtering by Kruskal Wallis test,the we use Gaussian Process Regression(GPR)to establish the conventional Radiomics model and obtain the Radiomics score.Habitat model was obtained by K-means clustering and with five clustering features and obtain the Habitat score.Finally,a Habitat_Radiomics combined model and Habitat_Radiomics score were obtained fitting the Radiomics score and Habitat score with the weighted 50%:50%.Receiver operating characteristic curve(ROC)was used for evaluation tne performance of the models.Delong test was performed for pairwise comparison of ROC curves.The 3 model scores above were combined with clinicopathological indicators(age,pretreatment tumor volume,FIGO stage,pretreatment SCC value,and presence of lymph node metastasis before treatment)and five clustering features(inertias,silhouette_score,CHI,DBI,and SP)in the habitat imaging model were used as candidate to construct multivariate Cox proportional risk model to obtain independent risk factors for PFS and OS,respectively.Then we use factors obtained above to divide the patients into high-risk and low-risk groups,and further assessed the differences by Kaplan-Meier survival curve through log-rank test to clarify the relationship between each independent predictor and PFS and OS.Results: Application of t-test,Mann-Whistney U-test and Chi-square test showed that the clinicopathological characteristics of patients randomly selected from the training and validation sets were equally distributed in the two data sets(p>0.05).In the habitat model,the optimal number of clusters K=3 was determined,i.e.,divided into three habitat subregions.The combined Habitat_Radiomics model of habitat imaging combined with conventional radiomics performed better and not worse than conventional radiomics model in the training set(AUC=0.944,P=0.026)and validation set(AUC=0.897,P=0.046).Clinical Decision Curve Analysis(DCA)showed that the Combined Habitat_Radiomics model outperformed conventional radiomics model in the training and validation sets in terms of net patient benefit.FIGO staging,inertias,SP and the scores of Habitat_Radiomics prediction model were independent risk factors for PFS;SP,FIGO staging and the scores of Habitat_Radiomics prediction model were independent risk factors for OS.Conclusion: The CE-MR-based combined model of habitat imaging and conventional radiomics can predict the early treatment response of concurrent chemoradiotherapy for locally advanced squamous carcinoma of the uterine cervix,while the score of the combined Habitat_Radiomics model is an independent prognostic risk factor for long-term prognosis PFS and OS.Therefore,it can provide a new method to assess the intratumoral spatial heterogeneity of LACC,which is expected to be a new imaging biomarker.Part II Exploring the Correlation between Physiological Intratumor Habitats and Spatial Tumor Subregions based on Magnetic Resonance Habitat Imaging by Xenograft Model of Cervical Squamous CarcinomaObjective: One of the important controversies in radiomics currently is the poor clinical interpretability,so the coregistered full-size pathological images with images provides a better way to validate and increase interpretability of radiomics.The aim of the second part of this study was to attempt to validate the correlation of MR-based habitat subregions with different physiological intratumoral subregions by establishing a xenograft model of cervical squamous carcinoma and obtaining pathological sections with immunohistochemistry(IHC)coregistered with MR-guided3 D printed tumor molds..Methods: Twelve xenograft models of cervical squamous carcinoma were successfully established by human cervical squamous carcinoma Siha cell line,and MR scans were performed separately.The ROI of tumors were manually segmented in T2-weighted images,respectively,and the spatial subregions of MR habitat imaging were confirmed by K-means clustering using T1 and T2 modalities.Then,we prepared histologic and IHC characterizations of sections coregistered using MR-guided 3D printed tumor molds and obtained 30 sections(including the largest area level and the adjacent upper and lower layers).Finally,these slices were performed with H&E and CD31,Ki-67 immunohistochemical staining using 3consecutive layers(4μm)coregistered with MR image layer.The physiological subregions were generated by agglomerative clustering after precise alignment using neighborhood analysis to form staining density maps respectively.At last,Pearson correlation analysis was performed on the area percentages of paired habitat imaging suregions and physiological intratumoral habitat.Results: Both MR images and histopathological slices were divided into three habitat subregions.A significant positive correlation was found between MR red subregion and high-cellularity(HC)histologic habitat,r~2 = 0.76(P < 0.01),a non-significant positive correlation was found between MR green subregion and high-vascularity(HV)histologic habitat,r~2 = 0.014(P = 0.042)and a non-significant positive correlation between MR blue subregion and non-HC and/or non-HV histologic habitat,r~2 = 0.61(P < 0.01).Conclusion: We successfully made the attemption to provide biological validation for spatial tumor subregions constructed by MR-based habitat imaging methods using a newer method of coregistering MR images and histologic sections.The physiological tumor habitats were roughly identified with MR-based habitat imaging in xenograft models of cervical squamous carcinoma.Meanwhile,the MR-based habitat imaging method can construct spatial tumor subregions and correlate with physiological tumor habitats..