Abstract

Background: Tumor-stroma ratio (TSR) has been recognized as a valuable prognostic indicator in various solid tumors. The aim of this study is to examine the clinicopathological relevance of TSR in Merkel cell carcinoma (MCC) using artificial intelligence (AI)-based parameterization of the stromal landscape. Methods: Virtual slides were partitioned into 128x128-pixel “mini-patches,” which were inputted into a novel classification framework termed TumOr And STromA (TOAST), whose output was the probability of the mini-patch representing tumor cells rather than stroma. A separate Counter-TOAST scheme was used to delineate regions of interest (ROI). The AI models were trained on six MCC samples and validated on four samples. For clinicopathological validation, the cohort was divided into train (n = 53) and test (n = 36) sets. Hierarchical random samplings of 50 mini-patches per region were performed throughout 50 regions per slide. TSR and landscape parameters were estimated by the maximum-likelihood algorithm. Results: Receiver Operating Characteristic curves showed that the areas under the curve (AUCs) of Counter-TOAST in discriminating ROI from space, hemorrhage, and necrosis were 1.00, 1.00, and 1.00, respectively. AUCs of TOAST in differentiating tumor cells from overall stroma, collagenous stroma, and lymphocytes in the test samples were 0.97, 1.00, and 0.94, respectively. AUCs of parameter-based models in predicting polyomavirus, nodal metastasis, distant metastasis, and recurrence were 0.69, 0.65, 0.74, and 0.82, respectively. Conclusions: TSR can be reliably calculated using an AI-based classification framework and can predict of various prognostic features of MCC. Larger studies are needed to confirm its clinicopathological significance.