We selected the venous phase of CT examination for analysis after comparing the ratio of ICC greater than 0.75 of radiomic features from two radiologists in three phases (72.3% of venous phase vs. 60.1% of unenhanced phase and 69.4% of arterial phase). Existing research has suggested that the venous phase was superior compared with arterial phase for lymph node assessment and the arterial phase was better for local tumor staging. We chose the venous phase to analyze the MSI status of RC patients. According the guidelines of National Comprehensive Cancer Network, patients with MSI status have a better prognosis and obtain no benefit from 5-FU-based adjuvant chemotherapy in stage II colorectal carcinoma. Therefore, it is significant to noninvasively and preoperatively predict the MSI status of RC patients. Recent developments in the field of radiomics have reported that combining radiomics with clinical factors could achieve a better predictive performance in predicting MSI status of patients with stage II colorectal carcinoma. While ,to best of our knowledge, there was no research focused on the tumoral and peritumoral radiomics to evaluate the MSI status of RC patients.
We conducted tumoral and peritumoral CT-based radiomics analysis and developed six machine learning algorithms to predict the MSI status of RC patients. We taken the indicator of RSD of 100 Bootstrap replication to assess the different performance of algorithms. The lower the RSD value of the algorithm was, the more stable its performance is. So we selected the algorithm of LR (RSD: 3.05%) to construct a integrative model of tumoral and peritumoral radiomics to assess the MSI status of RC patients compared with the algorithms of Bayes (RSD: 3.10%), SVM (RSD: 27.95%), RF (RSD: 7.26%), KNN (RSD: 4.68%), and DT (RSD: 8.20%). After the dimension reduction of radiomic features, there were 51 radiomic features remained to construct the M-LR. The AUCs of M-LR were 0.817 (95%CI, 0.772–0.856) in the training set and 0.726 (95%CI, 0.648–0.796). The MRI-based radiomics and machine learning showed that the Bayes-based radiomics signature performed better compared with other LR-based, SVM-based, KNN-based, and RF-based radiomics signature to predict the extramural venous invasion in RC patients. The deep learning based on high-resolution T2-weighted magnetic resonance images showed a good predictive performance for MSI status in RC patients. The multivariate analysis of previous study to predict the treatment response of RC patients found that RF and KNN achieved the highest AUC among pre-treatment and post-treatment features. In our study, the algorithm of LR with the minimal RSD showed the best performance in predict the MSI status of RC patients.
Previous study indicated that colorectal carcinoma with MSI status have distinct clinicopathological and pathological characteristics compared with these with MSS status, including proximal colon predominance, poor differentiation, and abundant tumor infiltrating lymphocytes. Therefore, we integrated the 51 selected radiomic features and significant clinicopathological variables of CEA, LNR, and drinking to construct a visual nomogram in predicting the MSI status of RC patients. The preoperative prediction of MSI status via CT-based radiomics adds specificity to clinical assessment and could contribute to personalized therapy. The radiomics nomogram incorporating radiomics signatures and clinical indicators of tumor location, patient age, high-density lipoprotein expression, and platelet counts could potentially be used to facilitate the individualized prediction of MSI status in patients with colorectal carcinoma. To best of our knowledge, there was no machine learning research incorporating CT-based radiomics and clinicopathological variables to predict the MSI status of RC. Our integrated radiomics-clinicopathological nomogram showed a better performance with AUCs of 0.843 (95%CI, 0.800–0.880) in the training set and 0.737 (95%CI, 0.659–0.805) in the validation set than the simple M-LR.
In conclusion, the present research explored, for the first time, the effects of CT-based tumoral and peritumoral radiomics with the machine learning algorithm of LR could help predict the MSI status of RC patients. Moreover, the visual radiomics-clinicopathological nomogram incorporating radiomics and significant clinicopathological variables of CEA, LNR, and drinking performed better prediction of MSI status of RC patients .