AI-based tumor-stroma ratio quantification algorithm: comprehensive evaluation of prognostic role in primary colorectal cancer

Tumor-stroma ratio (TSR quantification in primary colorectal cancer is an important prognostic parameter, with stroma-rich tumors considered to have an unfavorable prognosis. Earlier studies involved relevant region selection and TSR quantification by human analysts, who are prone to interobserver variability. The aim of the current study was to develop a fully automated, quantitative algorithm for TSR analysis based on precise segmentation of H&E-stained histological tissue sections. The TSR quantification algorithm was developed based on the segmentation backbone, allowing accurate pixel-wise mapping of all relevant tissue classes (n = 12), including tumor cells, tumoral stroma, necrosis, and mucin. Three well-characterized cohorts of patients with stage I–IV primary operable colorectal cancer and available digital H&E histological slides were included (n = 548, n = 147, and n = 622, respectively). Three sizes of area for TSR analysis were tested (1.0, 1.5, and 2.0 mm). The maximal TSR value per case was used for prognostic analysis involving different clinical endpoints. Regional heterogeneity of TSR was high in most tumors, with the algorithm effectively finding the most relevant region for analysis. Maximal case-level TSR values depended on the size of the area for analysis, which also significantly influences the prognostic performance of the TSR and must be a matter of standardization. In Cox analysis, an analytical size of 1 mm allowed the best performance, with an independent prognostic role retained in the context of other pathological variables for progression-free survival, cancer-specific survival, and overall survival endpoints. A powerful, fully automated, fully quantitative, objective tool for TSR assessment in primary colorectal cancer was developed and validated to have independent prognostic value. Standardization of TSR quantification is important given that analytical parameters can substantially influence the prognostic performance.