Study On Factors Influencing Tumor Regression Degree In Locally Advanced Rectal Cancer Following Neoadjuvant Chemoradiotherapy And Development Of Clinical Assessment Methods For Tumor Regression

Background:The epidemiology of colorectal cancer in China reveals that the incidence of rectal cancer is higher than that of colon cancer,with a significant number of newly diagnosed cases of locally advanced rectal cancer（LARC）.Due to the advanced tumor staging and complex surrounding anatomy of LARC,it profoundly impacts the long-term prognosis and quality of life of patients.Neoadjuvant chemoradiotherapy（nCRT）constitutes a key element of the standard treatment protocol for locally advanced rectal cancer（LARC）.Its application contributes to tumor volume reduction,facilitating enhanced preservation of the anus and rectum while simultaneously decreasing the local recurrence rate.Nonetheless,a considerable proportion of patients encounter suboptimal tumor regression or even disease progression,while also enduring adverse reactions to nCRT.At present,clinical practice lacks reliable methodology for predicting tumor regression outcomes following nCRT.On the other hand,approximately 11.4%to 27.0%of patients achieve complete tumor regression after nCRT;however,the absence of reliable evaluation methods for tumor regression in clinical practice results in numerous patients undergoing unwarranted surgical interventions.Hence,this study endeavors to identify the determinants influencing the degree of tumor regression following nCRT by examining the clinicopathological,radiomic,and molecular features of LARC patients treated with nCRT at our institution,and to develop a novel approach for the clinical assessment of tumor regression after nCRT,facilitating precise identification of the beneficiary population from nCRT and accurate execution of surgical interventions following nCRT.Method:This study is structured into four parts.Part Ⅰ:A retrospective cohort study was conducted utilizing clinicopathological data from patients with LARC who underwent nCRT at our institution between January 1^st,2011,and December 31^st,2021.The final cohort was established by applying inclusion and exclusion criteria.Patients were subsequently categorized into two groups according to the degree of tumor regression:the sustained complete regression（SCR）group and the non-SCR group.Clinicopathological characteristics between the two groups were compared using t-tests,Mann-Whitney U tests,chi-square tests,and Fisher’s exact probability tests.Logistic regression analysis was used to identify independent factors associated with achieving SCR.The predictive value of clinical indicators for tumor regression was assessed using the area under the curve（AUC）of the receiver operating characteristic（ROC）curve.Part Ⅱ:T2-weighted magnetic resonance imaging（MRI-T2WI）sequences before nCRT were retrospectively acquired for a subset of patients included in Part I of the study.Regions of interest（ROI）for the lesions were manually delineated,and three-dimensional volumes of interest（VOI）were generated using specialized software to extract radiomic features.Features exhibiting high consistency were retained for further analysis.Machine learning algorithms,such as logistic regression,decision tree,random forest,K-nearest neighbor,and support vector machine,were employed to develop predictive classifiers for predicting the degree of tumor regression after nCRT and determining treatment sensitivity or resistance.Nested resampling techniques were utilized for parameter optimization and performance evaluation during the construction of these classifiers.The ROC curve analysis was performed to assess the predictive performance of the classifiers.Part Ⅲ:Tumor tissue and peripheral blood cell samples were collected from LARC patients before nCRT for genomic and proteomic feature acquisition.In the genomic analysis,the genomic features of patients with different degrees of tumor regression were compared.In the proteomic analysis,patients were stratified into training and testing cohorts based on their enrollment time.In the training cohort,consensus clustering techniques based on proteomic features were utilized to classify tumors into distinct molecular subtypes.Differentially expressed proteins associated with tumor regression degree in each subtype were used to build predictive models for tumor regression degree in each molecular subtype.Finally,proteomic data from the testing cohort were applied to assess the predictive performance of model in each molecular subtype using ROC curve analysis.Part Ⅳ:Endoscopic and MRI data of LARC patients after nCRT but before surgery were retrospectively collected.The diagnostic performance of endoscopic features and the MRI-based tumor regression grade（mr TRG）system for pathological complete regression（p CR）were evaluated.A new MRI-based split scar sign（mr SSS）scoring system was developed based on morphological characteristics on MRI-T2WI and diffusion-weighted imaging（DWI）sequences,and its interobserver consistency and diagnostic performance for p CR were evaluated.Results:Part Ⅰ:A total of 732 patients were included,with 169 in the SCR group and 563 in the non-SCR group.There were statistically significant differences between the two groups in terms of clinical T stage,clinical lymph modes metastases status,extramural vascular invasion status,mesorectal fascia invasion status,and pre-treatment CEA levels.Pre-treatment CEA levels had predictive value for the incidence of SCR,with an AUC value of0.639（95%CI 0.571～0.706）.In the multivariate analysis,pre-treatment CEA≥3.36 ng/m L was an independent factor for the decreased incidence of SCR（OR=0.238,95%CI0.121～0.467,P<0.001）.There were no statistically significant differences in the routine blood test indicators and derived indices between the SCR group and the non-SCR group,and they had no predictive value for the occurrence of SCR.Waiting time affected the p CR rate after nCRT.When the waiting time was within period of 18～19 weeks,the proportion of p CR patients was the highest,reaching 44.4%;and within the period of 9～10 weeks,the cumulative p CR rate increased most rapidly.As the number of consolidation chemotherapy sessions increased after nCRT,p CR rates increased accordingly,and this trend had statistical significance（P=0.027）.Part Ⅱ:A total of 235 patients were included,with 104 patients in the nCRT-sensitive group and 131 patients in the resistant group.1158 radiomic features showed good consistency,and a total of 78 features were involved in the construction of the classifiers after reducing the collinearity.Five machine learning algorithms including logistic regression,decision tree,random forest,K-nearest neighbor,and support vector machine were used to establish classifiers predicting the degree of tumor regression after nCRT.The AUC values in the training sets ranged from 0.751 to 0.832,while those in the test sets ranged from 0.626 to 0.818.Part Ⅲ:A total of 144 patients underwent proteomic data collection,of which 66patients underwent genomic feature collection.There were 68 patients in the nCRT-sensitive group and 76 patients in the resistant group.There were no statistically significant differences in gene mutation rates of TP53,APC,KRAS,and BRAF between the two groups（P>0.05）,and tumor mutation burden（TMB）had no statistically significant difference（P>0.05）.There were statistically significant differences in the mutation rates of HERC2,BPTF,ADAMTS19,and RYR2 genes（P<0.05）.Based on proteomic features,100 patients in the training cohort were classified into three molecular subtypes,and the activated molecular pathways in the sensitive and resistant groups were different in each subtype.The models based on differentially expressed proteins in the three subtypes had AUC values of0.922,0.898,and 0.900 in the training cohorts,and AUC values of 0.762,0.754,and 0.643in the testing cohorts.Part Ⅳ:A total of 483 patients were included,with 84 p CR patients.The sensitivity,specificity,and accuracy of diagnosing p CR based on endoscopic features of normal mucosa or flat scars or apillary dilation were 0.333,0.919,and 0.810,respectively.The sensitivity,specificity,and accuracy of diagnosing p CR based on mr TRG=1 was 0.244,0.919,and 0.772,respectively.The mr SSS scoring system established based on MRI-T2WI and DWI sequences showed good interobserver consistency with a Kendall’s W coefficient of 0.899.The sensitivity,specificity,and accuracy of using an mr SSS score of 0 to diagnose p CR are0.667,0.974,and 0.902,respectively.Conclusion:1.Patients with pre-treatment CEA≥3.36 ng/m L have a lower rate of complete tumor response after nCRT;prolonging the waiting time after nCRT can increase the incidence of p CR,and adding several cycles of consolidation chemotherapy after nCRT can improve the occurrence of p CR.2.It is practical to develop classifiers using five machine learning algorithms including logistic regression,decision tree,random forest,k-nearest neighbors,and support vector machine based on pre-nCRT radiomic features,to predict the degree of tumor regression.Radiomic features have predictive value for tumor regression after nCRT and can be helpful in accurate identification of nCRT beneficiary population.3.There is a statistically significant difference in the mutation rates of the HERC2,BPTF,ADAMTS19,and RYR2 genes between nCRT-sensitive and resistant patients.Based on proteomic features,LARC can be molecularly subtyped,and within each subtype,a predictive model for tumor regression degree based on differentially expressed proteins has been built and tested,highlighting the important value of genomic and proteomic features in predicting tumor regression after nCRT.4.Endoscopic features and mr TRG have poor performance in assessing the degree of tumor regression after nCRT.The newly established mr SSS scoring system based on MRI-T2WI and DWI sequences has good interobserver consistency and diagnostic performance for p CR assessment,which can assist in making clinical decisions after nCRT.