Application analysis of radiomics models based on high-resolution T2-weighted MR images for evaluating the effectiveness of neoadjuvant chemoradiotherapy in locally advanced rectal cancer

Abstract

[Objectives] To explore the efficacy of radiomics models based on high-resolution T2-weighted MR images (HRT2WI) in evaluating the effectiveness of neoadjuvant chemoradiotherapy (nCRT) in locally advanced rectal cancer (LARC). [Methods] We retrospectively collected clinical data of patients with LARC who were treated at the First Affiliated Hospital of Naval Medical University from January to December 2018. The patients received nCRT (long-term radiotherapy combined with concurrent chemotherapy) and were confirmed to have rectal adenocarcinoma by postoperative pathology. High-resolution T2-weighted MR imaging of the rectum was performed in baseline status and after nCRT. Subjective evaluations of the effectiveness of nCRT was performed by radiologists based on the presence of split scar sign (SSS) on HRT2WI after nCRT. The postoperative pathological tumor regression grade (TRG) was used as the "gold standard" for evaluating the effectiveness of nCRT (good or poor outcomes). Region of interest was delineated on HR-T2WI in baseline status and after nCRT, and a volume of interest (VOI) was automatically generated. Radiomics features were extracted in two ways and corresponding models were built: Model 1 extracted radiomics features from baseline VOI; Model 2 extracted radiomics features from VOI after nCRT. Features with an intraclass correlation coefficient ≥ 0.8 were selected, and the least absolute shrinkage and selection operator were used for dimensionality reduction to select the optimal relevant radiomics features for evaluating TRG. Seventy percent of the cases were randomly selected as the training set to construct the XGBoost models (Model 1 and Model 2), and the remaining 30% of the cases were used as the test set to validate the two models separately. Receiver operating characteristic curves were plotted and analyzed for the results of the radiomics models and subjective evaluations of SSS [area under the curve (AUC), 95% confidence interval (CI) of the AUC, sensitivity, specificity, accuracy, precision, F1 score). The Delong test and calculation of the net reclassification index(NRI) were used to compare differences between models. Decision curve analysis (DCA) was used to assess clinical benefits. [Results] The clinical data of 189 patients were included in the analysis. The median time interval between the end of nCRT and surgical treatment was 68 (63, 74) days, and the median time interval between high-resolution T2-weighted MR imaging of the rectum after nCRT and surgical treatment was 6 (4, 9) days. Forty-one patients achieved pathological complete response, and 93 (49.2%) patients were evaluated as having good outcomes based on the postoperative pathological TRG. Eight and six optimal relevant radiomics features were obtained for Model 1 and Model 2, respectively. In the test set, the AUC for Model 1 and Model 2 were 0.977 (95%CI: 0.955-1.000) and 0.921 (95%CI: 0.865-0.978), respectively, while the AUC for subjective evaluations of SSS was 0.633 (95%CI: 0.540-0.727). The Delong test showed that compared with Model 1, P＝0.186 for Model 2, and the NRI for Model 2 compared with Model 1 was -0.402. The Delong test also showed that subjective evaluations of SSS was significantly different from Model 1 and Model 2, with P values less than 0.05, and the NRI for subjective evaluations of SSS compared with Model 1 and Model 2 was negative. The results of DCA indicate that within the probability threshold range of 0 to 1, the clinical benefits of Model 1 and Model 2 are superior to considering all patients as having good or poor outcomes, as well as superior to subjective evaluations of SSS. Furthermore, the clinical benefits of Model 1 are superior to that of Model 2. [Conclusion] Both the radiomics models based on HR-T2WI in baseline status and after nCRT can effectively evaluate the degree of tumor regression in LARC, with the radiomics model based on baseline images exhibiting relatively better performance.

Keywords

• locally advanced rectal cancer
• neoadjuvant chemoradiotherapy
• high-resolution t2-weighted mr images
• radiomics
• machine learning