Purpose

To predict pathological complete response (pCR) of locally advanced rectal cancer before neoadjuvant chemoradiotherapy (NCRT), by the combination of non-Gaussian diffusion kurtosis model (DKI) and radiomics model.

Method

474 patients were averagely divided into the discovery cohort (n=240) and the validation cohort (n=234) in chronological order. Surgically resected specimens were histopathologically examined as the reference. All MRI examinations were performed with a 3.0 T MR scanner (Discovery 750; GE Healthcare) using an 8-channel phased array torso coil within one week before the initiation of NCRT. DWI data were obtained by using SSEPI with 12 b-factors (0, 20, 50, 100, 200, 400, 600, 800, 1000, 1200, 1400 and 1600 s/mm²). TR=2800 ms, TE=70 ms, FOV=340×340 mm, Matrix=256×256, thickness=4.0 mm, gap=1.0 mm. D_app and K_app maps were fitted by DKI fitting algorithm ^[1]. Beside conventional parameteric maps D_app and K_app, S_app=log(S(800s/mm²)) is also used to include the structual information. The regions of interest (ROIs) of rectal tumor were manually created by radiologists on DWI slices with itk-SNAP (Fig.1). 59 quantitative imaging features were extracted from each parameteric map by homemade program on the platform of MATLAB (2011b, MathWorks). It includes 6 gray-level statistics features, 3 histogram features from 3d volumetric ROI, 3 histogram features, 8 GLCM features, 13 GLRLM features, 13 GLSZM features, 5 NGTDM features and 8 geometric features from the largest cross section. Feature selection was performed by logistic regression with best subset selection (BeSS) based on the training data defined by the discovery cohort with the BeSS package in R (R Core Team, 2017) recently developed by Wen, et al ^[2].

Result

Predictive model is first established with 59 features from S_app, D_app and K_app map individually (Fig.2). The ROC curves (Fig.3) show that S_app model (AUC 0.963,0.934) performs better than using D_app model (AUC 0.900, 0.852) or K_app model (AUC 0.924, 0.864) in both discovery (first value) and validation (second value) cohorts. If the features S_app, D_app and K_app are combined together (177 features), the resulting model shows improvement in both discovery (AUC=0.979) and validation cohorts (AUC=0.943). The model prediction performance by BeSS is compared with LASSO using R:glmnet package. Using the discovery cohort with 177 combined features, LASSO with cross validation selection of regularization parameter is repeated for 100 times, each time being assessed by the AUC for both discovery and validation cohorts. In Fig.4, the AUC scores from LASSO show variations among repeated runs, and their best scores are inferior to the combined model we have developed above using the BeSS method.

Discussion

NCRT is the standard treatment for locally advanced rectal cancer^[3]. Patients with pCR after NCRT could benefit from less invasive surgery or a “wait-and-see” strategy^[4]. However, it is still a challenge to accurately predict pCR preoperatively. Recently, the predictive accuracy of pCR is further improved by non-Gaussian DKI model and radiomics model from two different approaches. But the combination of the two models are rarely studied. Previous DKI research mainly analyzed the mean value inside the tumor ROI. However, more and more studies have confirmed that the heterogeneity in the tumor is related to the biological behavior. Therefore, the combination of DKI and radiomics provide more information. For comparision, mean D_app produces AUC=0.729 for discovery cohort and 0.770 for validation cohort and mean K_app produces AUC=0.572 for discovery cohort and 0.695 for validation cohort. They are significantly lower than that from 59 features together. Results show that S_app produces the highest AUC. Unlike D_app or K_app that only contains diffusion information, S_app is the logarithm form of diffusion and T2 weighted image. Therefore, the combination of structral and functional information produces higher predictive power. Previous popularly used LASSO method is known to have several shortcomings in certain scenarios such that it tends to bring in a large number of unimportant features and therefore results in biased coefficient estimates^[5]. To overcome such shortcomings of the lasso, the best subset selection (BeSS) approach is recommended in statistical modeling community^[6], even though it has been a rather challenging problem in computation^[7]. In this study we take the BeSS approach to build a logistic type of radiomics model by using the new R:BeSS package. It is justified to have superior prediction performance than the lasso method.

Conclusion

We developed a DKI based radiomics model with excellent performance for individualized, non-invasive prediction of pCR. This model may be used to identify LARC patients who can omit surgery after chemoradiotherapy.

Acknowledgements

No acknowledgement found.