Session Item

To study magnetic resonance imaging (MRI) based radiomic features for predicting tumor response to neoadjuvant chemoradiation (NACTRT) in patients of rectal cancers.

We retrospectively studied 100 patients (male:76, female:24) of rectal cancer, who underwent baseline and post-treatment MRI scans 6 weeks after therapy. The treatment protocol consists of NACTRT with concurrent capecitabine. The study was conducted after due approval of institutional ethics committee. Region of interest (ROI) was delineated along the tumor outline in all cross sections of baseline and post-treatment axial T2W MRI imaging. . First order texture (radiomic) features were extracted were extracted and filtered across various spatial scale filters (SSF 0-6) for quantification of histogram derived parameters, namely mean, standard deviation, entropy, mean positive pixel (MPP), skewness, and kurtosis. After NACTRT, 74 patients underwent complete surgical resection and their pathological specimen served as the gold standard for assessing pathological response Remaining 26 patients did not undergo surgery of which 5 were observed as part of wait and watch protocol and 21 were deemed unresectable due to local progression or distant metastasis. Receiving operating characteristic (ROC) curves were generated to distinguish between Complete response (CR) and partial response (PR) + no response (NR) with respect to values of individual texture features. Area under curve (AUC) and metrics such as sensitivity and specificity were used as measures of diagnostic accuracy.

In the entire cohort of 100 patients, 9 patients achieved radiological CR,  whereas 79 had PR and 12 showed NR. Pathologically 22 (29%) achieved  CR, 45 (60%) had PR and 7 (9%) showed NR. Among radiomic features of pretreatment scans that could best predict radiological response for the entire cohort, Skewness (SSF-3)  (AUC-0.821, 55.5% sensitivity, 96.7% specificity, 93% accuracy) was most predictive of response. On post-treatment scans, MPP (SSF 5) was the best predictor of response (AUC 0.907, 77.7% sensitivity, 92.31% specificity, 91% accuracy). For the 74 operated patients, pathological response  was best predicted by Skewness (SSF-2) (AUC = 0.721, 66.6% sensitivity , 73% specificity and 70.89% accuracy) on pretreatment scans. While kurtosis (SSF-6) was the best predictor of pathological response on post-treatment MRI scans. (AUC 0.663, 81.4% sensitivity , 46.15% specificity, 58.23% accuracy).

Tumor radiomics predicted for response to NACTRT with a fair degree of accuracy.  These results need to be further validated in prospectively conducted studies including a larger cohort of patients, if clinical decisions are to be guided by radiomics in future.