MRI-based radiomic score increased mrTRG accuracy in predicting rectal cancer response to neoadjuvant therapy

Abstract

Purpose: To develop a magnetic resonance imaging (MRI)-based radiomics score, i.e., "rad-score," and to investigate the performance of rad-score alone and combined with mrTRG in predicting pathologic complete response (pCR) in patients with locally advanced rectal cancer following neoadjuvant chemoradiation therapy.

Methods: This retrospective study included consecutive patients with LARC who underwent neoadjuvant chemoradiotherapy followed by surgery from between July 2011 to November 2015. Volumes of interest of the entire tumor on baseline rectal MRI and of the tumor bed on restaging rectal MRI were manually segmented on T2-weighted images. The radiologist also provided the ymrTRG score on the restaging MRI. Radiomic score (rad-score) was calculated and optimal cut-off points for both mrTRG and rad-score to predict pCR were selected using Youden's J statistic.

Results: Of 180 patients (mean age = 63 years; 60% men), 33/180 (18%) achieved pCR. High rad-score (> - 1.49) yielded an area under the curve (AUC) of 0.758, comparable to ymrTRG 1-2 which yielded an AUC of 0.759. The combination of high rad-score and ymrTRG 1-2 yielded a significantly higher AUC of 0.836 compared with ymrTRG 1-2 and high rad-score alone (p < 0.001). A logistic regression model incorporating both high rad-score and mrTRG 1-2 was built to calculate adjusted odds ratios for pCR, which was 4.85 (p < 0.001).

Conclusion: Our study demonstrates that a rectal restaging MRI-based rad-score had comparable diagnostic performance to ymrTRG. Moreover, the combined rad-score and ymrTRG model yielded a significant better diagnostic performance for predicting pCR.

Keywords: Magnetic resonance imaging; Neoadjuvant therapy; Radiomics; Rectal cancer.

Fig. 1

Flowchart demonstrating patient inclusion into the study Abbreviations: CRT: chemoradiation therapy; LARC: locally advanced rectal cancer; MRI: magnetic resonance imaging.

Fig. 2

Example of tumor and tumor bed segmentation. 53-year-old woman with locally advanced rectal cancer staged as cT3N0 at baseline. Volumes of interest (VOIs) were manually segmented of the tumor on baseline rectal MRI T2WI (A) and of the tumor bed on restaging rectal MRI T2WI (B). Volume rendering of the segmentations on baseline (C) and restaging (D) rectal MRI T2WI

Fig. 3

Manhattan plot of all extracted baseline MRI T2WI (A) and restaging MRI T2WI (B) radiomic features. The Manhattan plot reflects the P-values for all radiomic features between pathologic complete response (pCR) (A) versus non-pCR (B) groups, and their distribution among the different radiomic classes. Radiomic features are lined up on the x axis, whereas the -log2 (P) values are plotted on the y axis. The dashed line indicates the Bonferroni-corrected P value of 0.0012. Radiomic parameters above the dashed line were considered statistically significant Abbreviations: FO: first order; GLCM: gray level co-occurrence matrix; GLDM: gray level dependence matrix, GLRLM: gray level run length matrix; GLSZM: gray level size zone matrix; NGTDM: neighborhood gray tone difference matrix.

Fig. 4

Waterfall charts demonstrating the difference of each patient’s rad-score from the optimal rad-score cut-off value of −1.49, for all patients with and without pathologic complete response (pCR) in the training (n = 125) (A) and testing (n = 55) (B) sets. (C–D) Violin plots showing in red the rad-scores of patients who did not show pCR and in blue the rad-scores of patients who had pCR in the training (C) and testing (D) sets, demonstrating that rad-score was able to identify most patients with pCR