Beyond a single ratio: an adaptable deep learning approach to analyze the prognostic value of TSP distribution in ovarian cancer

The tumor-stroma percentage (TSP) has been proposed as a prognostic factor in ovarian cancer. When assessed manually by pathologists, the TSP can be subject to significant inter-rater variability. Automated methods for TSP determination have recently been developed. Existing approaches typically yield a single TSP value for a relatively small region of interest in the slide, introducing bias through region selection and potentially overlooking the TSP distribution across the entire histologically examined area. We present an adaptable deep learning pipeline, efficiently fine-tuned for ovarian cancer tissue analysis using a few-shot learning approach, requiring only minimal new training annotations. Instead of calculating a single TSP value, our method quantifies the global TSP distribution across the entire tumor and therefore captures its spatial heterogeneity. To identify a prognostic biomarker, we systematically evaluated a panel of metrics describing this distribution in 98 patients. Our analysis revealed that the mean-median difference of the TSP distribution [hazard ratio (HR) = 0.60, 95% CI = 0.46-0.79, p < 0.001], a quantitative measure of asymmetry, together with the mean TSP (HR = 0.65, 95% CI = 0.43-0.97, p = 0.03), are biomarkers significantly associated with recurrence-free survival. By successfully adapting a pipeline for ovarian cancer, we demonstrate that automated analysis of the entire TSP distribution provides significant prognostic information superior to that obtained by manual assessment. Spatial heterogeneity of the TSP is a key prognostic factor.