Automated determination of tumor cell percentages in whole slide images: A nuclear classification study for molecular pathology tests

Calculation of tumor cell percentage, a critical pre-analytical component in molecular pathology, is typically performed by pathologists estimating a ratio. This semiquantitative approach can lead to inter-observer variability, potentially adversely affecting patient management and treatment outcomes. In era of digital pathology, it became crucial to automate such assessments for more objective approach. This study aims to contribute to this process by developing a model for automated calculation of tumor cell percentage in high-grade serous carcinomas. Tumor containing hematoxylin-eosin slides from 100 patients were divided into training, validation, and test groups. Slides were digitalized and placed in QuPath platform. Image patches were obtained from WSIs of training and validation sets, and were stitched together to form digital microarrays by using ImageJ extension. Subsequently, nuclear detection and segmentation were performed using StarDist software, and tumor and non-tumor cell nuclei were classified using annotations. For binary classifier, random forest algorithm was selected. With hyperparameter tuning, many pre-models were assessed by cross-validation and most suitable pre-model was selected to apply to test set. Testing was performed on WSIs and criterion standard was based on corresponding immunohistochemistry (p53 or PAX8) slides which showed diffuse positivitity for tumor cells. Performance of model was measured using regression metrics. This study is designed to perform and assess a classifier in whole slide images to reflect real-world experience.