CT radiomics identifying non‐responders to neoadjuvant chemoradiotherapy among patients with locally advanced rectal cancer

Background and Purpose Early detection of non‐response to neoadjuvant chemoradiotherapy (nCRT) for locally advanced colorectal cancer (LARC) remains challenging. We aimed to assess whether pretreatment radiotherapy planning computed tomography (CT) radiomics could distinguish the patients with no response or no downstaging after nCRT from those with response and downstaging after nCRT. Materials and Methods Patients with LARC who were treated with nCRT were retrospectively enrolled between March 2009 and March 2019. Traditional radiological characteristics were analyzed by visual inspection and radiomic features were analyzed through computational methods from the pretreatment radiotherapy planning CT images. Differentiation models were constructed using radiomic methods and clinicopathological characteristics for predicting non‐response to nCRT. Model performance was assessed for classification efficiency, calibration, discrimination, and clinical application. Results This study enrolled a total of 215 patients, including 151 patients in the training cohort (50 non‐responders and 101 responders) and 64 patients in the validation cohort (21 non‐responders and 43 responders). For predicting non‐response, the model constructed with an ensemble machine learning method had higher performance with area under the curve (AUC) values of 0.92 and 0.89 as compared to the model constructed with the logistic regression method (AUC: 0.72 and 0.71 for the training and validation cohorts, respectively). Both decision curve and calibration curve analyses confirmed that the ensemble machine learning model had higher prediction performance. Conclusion Pretreatment CT radiomics achieved satisfying performance in predicting non‐response to nCRT and could be helpful to assist in treatment planning for patients with LARC. Non‐invasive methods to predict the response to neoadjuvant chemoradiotherapy in patients with locally advanced rectal cancer could help clinicians develop individualized treatment strategies. In this work, we retrospectively enrolled 215 patients with locally advanced rectal cancer who received neoadjuvant chemoradiation therapy, and we analyzed their pretreatment radiotherapy planning CT images for both radiomic and traditional radiological features. A machine learning model built with the selected radiomics and clinicopathological features. achieved satisfying performance for predicting non‐response to neoadjuvant chemoradiation therapy with an area under t