Artificial Intelligence Imaging Predictors for Rectal Cancer Management

ABSTRACT: Colorectal cancers are the third most frequently occurring cancer in the military, occurring in up to 8% of Veterans. Veteran rectal cancer patients are often diagnosed earlier relative to the general population, but

do not have improved survival rates; indicating limitations of the existing “one-size-fits-all” therapeutic protocol. The lack of personalized therapy in rectal cancers disproportionately impacts Veteran colorectal cancer patients who tend to be older (median age 67.5-years) and often have a higher comorbidity burden than the general U.S.

population. A major clinical question thus remains selection of an appropriate combination of neoadjuvant therapy, chemotherapy, and surgery to ensure optimal patient outcomes for Veteran rectal cancer patients. For instance, total neoadjuvant therapy (neoadjuvant chemotherapy and chemoradiation) can enable ~30% of

rectal cancer patients to achieve a pathologic complete response (no residual tumor). These patients are ideal candidates for organ-preserving strategies and non-operative management, such as a “watch-and-wait” (W&W) protocol. Conversely, up to 20% of patients may not exhibit any pathologic regression after neoadjuvant therapy,

but instead suffer tumor invasion to surrounding structures (lymphatic, vascular, or perineural) and thus a higher chance of life-threatening distant metastasis. Both of these issues particularly impact Veteran rectal cancer patients due their increased age, additional financial and psychological stresses, and survival benefits when

administered targeted adjuvant therapy to combat metastasis. Unfortunately, there is a paucity of clinically validated markers predictive or diagnostic of response to chemoradiation in rectal cancers and MR imaging has limited sensitivity/specificity for detecting treatment response or mutational/invasion status. This suggests a

critical need for accurate and objective non-invasive markers to (a) identify Veteran rectal cancer patients who have achieved complete response after neoadjuvant therapy so they can be safely recommended W&W, as well as (b) predict which Veteran patients will see added benefit from chemotherapy beyond chemoradiation alone.

Our initial set of novel artificial intelligence-based descriptors (known as radiomics) that capture heterogeneity, morphometry, as well as specialized measurements of lesion complexity on MRI have shown significant success for (a) distinguishing pathologic responders to chemoradiation with up to AUC=0.86, (b) predicting high-risk

metastatic rectal tumors with up to AUC=0.81, as well as (c) capturing heterogeneity in pathologic tissue organization associated with response and metastasis well as mutation status (N>200 patients, 3 institutions including the Northeast Ohio VA). In this VA Merit Award, we propose to develop and optimize our AI radiomic

descriptors to build clinically actionable computational image Rectal Response Classifier (ciRRC) tools to help (a) proactively select high-risk Veterans who will gain additional benefit from targeted chemotherapy based on predicted chemoradiation resistance or tumor invasion, and (b) identify Veteran rectal cancer patients with

minimal/dying tumor regions who can be safely recommended non-operative organ-preserving management. In Aim 1, we will optimize our specialized heterogeneity, morphometry, and lesion complexity radiomic features to capture imaging signatures from pathologically mapped tumor habitats on MRI (via spatial alignment of ex

vivo digitized whole-mount specimens with MRI) and thus distinguish pathologic tumor response of rectal tumors using a multi-institutional cohort of rectal cancer patients. In Aim 2, we will develop a ciRRC-Dx model for identifying patients at high-risk for distant metastasis or non-response and thus will see added benefit from

adjuvant chemotherapy; using baseline MRI scans from retrospectively accrued Veteran rectal cancer patients. Additionally, we will evaluate biological associations of ciRRC features with mutation status (of KRAS, BRAF, and NRAS genes) as well as physiological invasion (lymph, vascular, perineural). In Aim 3, we will develop a

ciRRC-Tx model for selecting complete responder patients who could be candidates for W&W, which will be validated on a multi-institutional cohort of Veteran rectal cancer patients from across Midwest VAHS centers.