Modeling Acute Chemoradiotherapy (CRT) Diarrhea Severity Using Automatically Contoured Small Bowel

Abstract

Acute severe diarrhea is a common toxicity in rectal cancer patients receiving CRT. A better understanding of the radiation tolerance of the small bowel is needed especially for novel radiation drug combinations. We investigate the dose impact by providing auto-contoured small bowel, using data from the ARISTOTLE phase III trial ISRCTN: 09351447. A subset of participants (n = 93/564) with locally advanced rectal cancer in the ARISTOTLE trial testing the addition of concurrent irinotecan (n = 48) to neoadjuvant capecitabine (n = 45) CRT (45/25 Gy/fx), in an MRI defined high risk of loco-regional failure. CRT was delivered with conformal techniques. Diarrhea was measured using CTCAE v4.0 weekly. We applied an AI-based auto-contouring model to segment the small bowel on planning CT. The small-bowel DVH parameters were combined with the treatment arm, age, sex and MRI-defined tumor stage in a linear regression (LR) model to predict acute diarrhea severity. Explainable Shapley values (conditional marginalized expectation of a machine learning model per feature) were used to quantify the independent and normalized impact of radiation dose vs Irinotecan on the likelihood of severe diarrhea. The auto-contouring model accuracy was consistent with clinical practice (mean dice coefficient = 0.739) and clinically acceptable when reviewed by clinicians. The treatment arm, MRI-defined T stage and small-bowel mean dose were found to be independently correlated to the diarrhea severity (p<0.001). V30Gy showed the strongest correlation to diarrhea severity in all the DVH parameters. The LR using the three variables yielded mean AUC scores of 0.898 (95% CI: [0.831,0.958]) on predicting Grade 2 and higher diarrhea, and 0.774 (95% CI: [0.678,0.869]) on predicting Grade 3 diarrhea based on 10-fold cross-validation. Shapley values showed that V30Gy>30.56 cm3 increases the likelihood of more severe diarrhea against the average (grade = 1.03) in the cohort. The impact of irradiation will be larger than the usage of Irinotecan within the patients with V30Gy >160.93 cm3. We accurately modelled acute diarrhea (AUC = 0.90) for rectal cancer patients receiving CRT using AI-contoured small-bowel structures. The treatment arm and small-bowel dose were independently correlated to the diarrhea severity. The explainable model allowed us to quantify the impact of radiation dose, usage of irinotecan, and its combination, with a threshold of V30Gy = 160.93 cm3 yielding an equivalent impact. We will be extending the analysis to the whole trial cohort to improve the statistical power.