Computerized features of tumor diversity on pre-chemoradiation MRI are associated with pathologic complete response in rectal cancers: A multi-institutional study.

Abstract

3608 Background: Neoadjuvant chemoradiotherapy (nCRT) prior to total mesorectal excision (TME) reduces the risk of local recurrence and is a standard-of-care treatment for patients with locally advanced rectal cancer. Previous studies have indicated up to 30% of patients who undergo TME demonstrate a pathologic complete response (pCR) to nCRT [2], but current serum markers and clinical evaluation have limited utility in predicting benefit from nCRT. In this study, we evaluated if tumor diversity features on baseline pre-treatment MRI were predictive of pCR in rectal cancers. Methods: Pre-nCRT T2-weighted MRIs from 126 patients who later underwent nCRT and TME (without neoadjuvant chemotherapy) were retrospectively collected from 2 institutions. 69 patients (14 pCR, Institution 1) formed the training cohort D1, while 57 patients (10 pCR, Institution 2) formed the independent holdout validation cohort D2. Response groupings were based on AJCC tumor regression grade (TRG) evaluation where pCR corresponded to no residual tumor cells in excised histopathologic specimens after nCRT (i.e. TRG0) and non/partial response otherwise (TRG1-3). Computerized tumor diversity algorithms extracted measurements of fractal dimensions, surface topology, and persistent homology from annotated tumor regions on MRI to quantify visual patterns, spatial geometry, and prolonged structural similarity. Top tumor diversity features identified via Wilcoxon rank-sum testing were used to train a random forest (RF) machine learning classifier to predict the likelihood of pCR using D1 and then independently evaluated for predicting pCR in D2. Results: Machine learning analysis identified 5 tumor diversity measurements associated with pCR on pre-nCRT MRI. The associated RF classifier for pCR achieved an AUC of 0.98 (95% CI, 0.9808-0.9949) in D1 and an AUC of 0.9117 in hold-out validation on D2. This corresponded to predicting 100% of non/partial response prior to nCRT as well as 80% of pCR patients via tumor diversity features on pre-nCRT MRI. Conclusions: Computerized tumor diversity features on pre-nCRT MRI can predict pathologic complete response to nCRT in rectal cancers. These findings need to be validated in larger multi-institutional cohorts to establish independent predictive utility of MRI-based biomarkers, which could aid in the selection of patients who will see maximal response from neoadjuvant therapy and avoid unnecessary surgery.