Deep Learning Model to Assist Pathological Complete Response Diagnosis in the Locally Advanced Rectal Cancer: A Prospective and Multicenter Study

Abstract

<h3>Purpose/Objective(s)</h3> Previously, through extracting post-neoadjuvant treatment magnetic resonance imaging (MRI) features, we constructed a radiomics based deep learning model to identify pathological complete response (pCR) patients from non-pCR LARC patients. Here, we aim to investigate the power of this deep learning system in a prospective and multicenter clinical study (Clinical Trial. gov No. NCT04278274). <h3>Materials/Methods</h3> In this study, patients who are pathologically diagnosed rectal adenocarcinoma and MRI defined as locally advanced stage (II and III stages) will be enrolled. All patients will be treated with neoadjuvant chemoradiotherapy or chemotherapy, following TME surgery and adjuvant chemotherapy. The images of MRI will be collected after the neoadjuvant treatment. Further, based on the post-neoadjuvant treatment MRI, nine radiologists divided into three different levels will be employed to generate the predicted response ("predicted as pCR" vs. "predicted as non-pCR"). The diagnosis from radiologists was compared with a reference diagnosis from the deep learning system. If two diagnoses did not match, the radiologists could then choose to adhere to their own diagnosis or adopt the diagnosis from the deep learning system as the final diagnosis. Both the initial and final assisted diagnoses were recorded. Finally, the pathologic report of TME surgery specimen ("confirmed pCR" vs. "confirmed non-pCR") will be served as a standard. The prediction accuracy, specificity, sensitivity as well as the area under curve (AUC) of Receiver Operating Characteristic (ROC) curves of identifying the pCR candidates from non-pCR individuals of the radiologists will be calculated. <h3>Results</h3> From February 2020 to July 2021, 205 patients were recruited in the prospective study. The preliminary predictive performance in deep learning system yielded a maximum AUC (0.850), which was higher than that of the radiologists (0.750, <i>p</i> < 0.01). Furthermore, the deep learning system improved the pooled AUC of the radiologists from 0.750 when diagnosing without deep learning system to 0.830 after supporting by deep learning system (<i>p</i> < 0.05). <h3>Conclusion</h3> Promisingly, based on the present prospective, multicenter clinical study, our deep learning prediction system displayed a higher predictive power to identify pCR patient using post-neoadjuvant treatment MRI. Importantly, this deep learning system may function as a diagnosis supporter to radiologists, particularly to the junior physician. Hopefully, the final and detailed result could be shared at 2022 ASTRO meeting.