Objective: With the increasing energy surrounding the development of artificial intelligence and machine learning (AI/ML) models, the use of the same external validation dataset by various developers allows for a direct comparison of model performance. Through our High Throughput Truthing project, we are creating a validation dataset for AI/ML models trained in the assessment of stromal tumor-infiltrating lymphocytes (sTILs) in triple negative breast cancer (TNBC).
Materials and methods: We obtained clinical metadata for hematoxylin and eosin-stained glass slides and corresponding scanned whole slide images (WSIs) of TNBC core biopsies from two US academic medical centers. We selected regions of interest (ROIs) from the WSIs to target regions with various tissue morphologies and sTILs densities. Given the selected ROIs, we implemented a hierarchical rank-sort method for case prioritization.
Results: We received 122 glass slides and clinical metadata on 105 unique patients with TNBC. All received cases were female, and the mean age was 63.44 years. 60% of all cases were White patients, and 38.1% were Black or African American. After case prioritization, the skewness of the sTILs density distribution improved from 0.60 to 0.46 with a corresponding increase in the entropy of the sTILs density bins from 1.20 to 1.24. We retained cases with less prevalent metadata elements.
Conclusion: This method allows us to prioritize underrepresented subgroups based on important clinical factors. In this manuscript, we discuss how we sourced the clinical metadata, selected ROIs, and developed our approach to prioritizing cases for inclusion in our pivotal study.
Keywords: Data; Prioritization; Sampling; Validation.