Multiparametric MR Images-Based Prediction Of Response To Neoadjuvant Therapy Of Rectal Cancer

Colorectal cancer ranks third in the global incidence rate and second in the mortality rate,ofwhich one third is rectal cancer,and most of them are locally advanced rectal cancer(LARC).The treatment for patients with LARC is total mesorectal excision(TME)following neoadjuvant therapy(NAT).The neoadjuvant therapies of LARC include neoadjuvant chemoradiotherapy and neoadjuvant chemotherapy(NCT).Neoadjuvant therapy can effectively reduce the risk of local recurrence and improve prognosis,but some patients show non-response to neoadjuvant therapy.On the one hand,accurately predicting non-responders before starting neoadjuvant therapy is important for avoiding overtreatment,and accurately predicting non-responders for NCT before surgery is important for choosing alternative treatments in time.On the other hand,lymph node status is a key factor for the recommendation of organ preservation for LARC patients following neoadjuvant therapy.But the efficacy evaluation of neoadjuvant therapy depends on the results of pathological specimens after surgery.Magnetic resonance imaging(MRI)is the most useful diagnostic imaging approach for rectal tumors.Radiomics can extract high-throughput features from images and use machine learning methods to build prediction models,which has potential in predicting the efficacy of NAT.This article focused on MRI-based radiomics for LARC and did three related work as follows:(1)Prediction of non-responders to neoadjuvant therapy in LARC using MRIbased radiomics.1)We pre-therapeutively predicted non-responders to neoadjuvant therapy in LARC.We retrospectively enrolled 425 patients with LARC in a single center.The multiparametric MRI-based radiomic model was developed using by MRI before neoadjuvant therapy,which achieved AUC of 0.773 in the validation set,and performed better than all single-modality models.2)We preoperatively predicted nonresponders to NCT in LARC.This retrospective study included 251 LARC patients in two centers.Pre-and post-NCT MRI data were collected,and radiomic features were extracted,including the changes in the radiomic features before and after NCT.The model predicting nonresponders was constructed using multivariate logistic regression analysis.The post+pre+diff model was established by radiomic features from pre-and post-NCT MRI images.In the multi-institution validation,the post+pre+diff model based on eleven features showed the best predictive performance,with AUCs of 0.912 and 0.895 in the internal cohort and external validation cohort.This study showed the proposed models have potential in predicting nonresponders to NAT before commencing this therapy,and potentially be used to preoperatively predict nonresponders to NCT in patients with LARC.(2)We preoperatively predicted lymph node status after neoadjuvant therapy in LARC using MRI-based radiomics.This study retrospectively enrolled 391 patients with LARC,and radiomics signature was developed using MRI after neoadjuvant therapy.Radiomics signature and assessment results(ymrT/N)of radiologist were included in multivariate analysis to build a combined model.The combined model predicted lymph node metastasis(LNM+)achieved AUC of 0.818 and a negative predictive value(NPV)of 93.7%in the validation cohort.Stratified analyses showed that the combined model could predict LNM+with the NPV of 100%in ymrT1-2 subgroup and achieves AUC of 0.957.This study reveals the combined model can accurately predict the negative lymph nodes for ymrT1-2 before surgery for patients with LARC following neoadjuvant therapy.(3)In order to replace the heavy segmentation task in radiomics,we developed and evaluated a deep learning-based automatic MRI rectal tumor segmentation method for prediction of non-responders to neoadjuvant therapy in LARC using radiomics.Dice of rectal tumor segmentation for T2WI before neoadjuvant therapy were 0.800 and 0.720 in the internal cohort and external validation cohort.Radiomics features extracted using manual and automatic segmentations agreed well(internal validation cohort:ICC=0.820,interquartile range:0.730-0.870;external validation cohort,ICC=0.750,interquartile range:0600-0.860).AUCs of radiomic prediction model using automatic segmentation had no statistical difference with that of the prediction model using manual segmentation in the internal(0.630 vs.0.690,p=0.176)and external(0.720 vs.0.770,p=0.101)validation cohorts.This study reveals the feasibility of replacing manual segmentation for T2WI before neoadjuvant therapy with automatic in the radiomics analysis.