3450 - Interpretable Artificial Intelligence to Predict Treatment Response in Rectal Cancer

X. L. Pang¹, X. B. Chen², L. L. Feng³, M. P. Hong⁴, P. Xie⁵, K. K. Wei⁶, J. Shi⁷, H. Chen⁸, F. He¹, Z. Y. Liu⁹, X. J. Fan¹⁰, and X. B. Wan¹¹; ¹Department of Radiation Oncology, the Sixth Affiliated Hospital of Sun Yat-sen University, Guangzhou, China, ²Guangdong Provincial Key Laboratory of Artificial Intelligence in Medical Image Analysis and Application, Guangzhou, China, ³Department of Radiology, Sun Yat-sen University Cancer Center, Guangzhou, China, ⁴Department of Radiology, JiaXing TCM Hospital Affiliated to ZheJiang Chinese Medical University, Jiaxing, China, ⁵Department of Radiology, the Sixth Affiliated Hospital of Sun Yat-sen University, Guangzhou, China, ⁶Biomedical Innovation Center, the Sixth Affiliated Hospital, Sun Yat-sen University, Guangzhou, China, ⁷GuangDong Provincial Key Laboratory of Colorectal and Pelvic Floor Diseases, The Sixth Affiliated Hospital, Sun Yat-sen University, Guangzhou, Guangdong, China, ⁸the Sixth Affiliated Hospital of Sun Yat-sen University, Guangzhou, China, ⁹Department of Radiology, Guangdong Provincial Peoples Hospital (Guangdong Academy of Medical Sciences), Southern Medical University, Guangzhou, China, ¹⁰Academy of Medical Sciences, Zhengzhou University, Zhengzhou, Henan, China, ¹¹Department of Pathology, the First Affiliated Hospital of Zhengzhou University, Zhengzhou, Henan, China

Purpose/Objective(s): Preoperative assessment of pathologic complete response (pCR) to neoadjuvant therapy (NAT) is an urgent need for organ preservation in patients with locally advanced rectal cancer (LARC). We previously built an artificial intelligence model named as RAPIDS which performed well in pCR prediction, but the ‘black box’ effect had hampers its clinical application. To improve the interpretability, we constructed an interpretable RAPIDS-II model combining with the SHapley Additive exPlanations analysis (SHAP) technique, which was trained on a large-scale dataset using the post-NAT magnetic resonance imaging (MRI) and clinicopathological factors. Here, we conducted a multicenter prospective study (NCT04278274) to further validated the performance and clinical applicability of RAPIDS-II model. Materials/

Methods: Patients with stage II/III rectal cancer and planning to receive TME surgery after neoadjuvant chemoradiotherapy were enrolled from two independent hospitals. Images of post-NAT MRI and clinicopathological information (clinical T stage, N stage, CEA level before and after NAT) were collected and inputted to RAPIDS-II model for pCR prediction. Meanwhile, six radiologists (divided into junior and senior groups) visually assessed the post-NAT MRI alone or with AI assistance by referring to the prediction results and SHAP visualization of RAPIDS-II model. The radiologists could choose to stick with their diagnosis or adopt the prediction from the RAPIDS-II model as their adjusted decision. Both the initial and final diagnoses of radiologists were recorded. All patients and physicians were blinded to the prediction results from either RAPIDS-II or radiologists, and the pathological report of the TME surgery specimen served as a standard. Area under curve was the primary outcome.
Results: Totally, 207 patients were recruited in the prospective study. RAPIDS-II performed robustly with an AUC of 0.795 (95%CI 0.723-0.859) in identifying the pCR patients. Importantly, RAPIDS-II assistance led to improvements in the overall AUC of junior radiologists from 0.673 (95%CI 0.605-0.741) to 0.758 (95%CI 0.691-0.825) and that of senior radiologists from 0.829 (95%CI 0.768-0.89) to 0.842(95%CI 0.738-0.902). Moreover, SHAP visualized how the radiomics-based Radscore of post-NAT MRI and clinicopathological factors impacted on RAPIDS-II output, where Radscore showed a positive correlation with pCR probability and was identified as the feature attributing most to RAPID-II performance.
Conclusion: The interpretable RAPIDS-II model performed well in pCR prediction, providing a reliable and useful tool in real world scenarios to tailor therapy for LARC patients. The final result will be announced at 2024 ASTRO meeting.