Tumour radiomics as imaging biomarker of tumour response to chemo-radiotherapy in patients with adenocarcinoma of rectum.

Abstract

243 Background: We aimed to study whether texture (Radiomic) features obtained from T2W MRI of patients with rectal cancer can be used as a surrogate imaging biomarker to predict response to NACTRT. We explored various machine learning tools to develop the best model to predict response to NACTRT. Methods: One hundred patients with stage II/III who underwent MRI before and after NACTRT and surgical treatment were enrolled. Patients were classified into complete response (pCR, n = 21) and partial and nonresponse (pPR + pNR, n = 53) on the basis of histopathological report (74 patients who underwent surgery) and clinico-radiological response (26 patients who did not undergo surgery). Tumor volumes (Region of interest) were manually selected in each tumour segment. There were sixty four first-order and higher-order radiomic features. Recursive feature elimination method was used for feature selection and 5 prediction models were tested using 10-fold cross validation for predicting tumour response to NACTRT. Results: Using prediction model assessment matrix (RFC, SVC, GBC, NBC, ABC), the best results for prediction response were obtained using the random forest model with AUC of 0.79 ± 0.15 (Mean ± Standard deviation) accuracy of 0.72 ±0.12, precision of 0.77 ± 0.10, sensitivity of 0.87 ± 0.07, f1 score- 0.81 ±0.07. By using random forest model, Texrad features obtained from pre and post NACTRT MRI could accurately predict pCR. To the best of our knowledge, apart for the added values of RF model to Texrad features, is the first study to demonstrate and compare RF, SVC, GBC, NBC, Adaboost models to the same cohort of patient with 10-fold validation. Conclusions: Our study shows that RFM was most accurate and stable for predicting tumour response. MRI based Radiomic features can be used as surrogate image biomarker to accurately predict the treatment response to NACTRT in patients with locally advanced rectal cancer. The prediction model can be used as a complementary non-invasive tool to identify patients eligible for organ preservation. [Table: see text]