Data harvesting vs data farming: A study of the importance of variation vs sample size in deep learning-based auto-segmentation for breast cancer patients

The aim of this study was to investigate the difference in output, when training a model in three different scenarios: a large clinical delineated data set (with 700/78 patients for training/testing, from the Danish Breast Cancer Group (DBCG) RT Nation Study), a clinical but curated dataset (with 328/36 patients for training/testing, from the DBCG RT Nation Study) and a smaller, but dedicated data set created by delineation experts (with 123/14 patients for training/testing, consensus delineations created by delineation experts). The model performance was estimated based on the performance metrics dice similarity coefficient (DSC), Hausdorff 95th percentile (HD95) and mean surface distance (MSD). Models were tested in test sets from their own cohort, and afterwards also compared in the dedicated data test set. The difference between model output was finally estimated by measuring the mean width and cranial caudal length of the model output for the models. When testing the model output between the clinical models and the dedicated models in their own test set, the two clinical models had a poorer performance, than the dedicated models, but not all metrics showed statistically significance. When testing the models in the dedicated data, the dedicated model showed a slightly better performance, along with fewer segmentation outliers. As a way of taking advantage of the strength from both types of data set, it could be an option to use a large clinical data set as a baseline model, and then finetune with smaller sized cohorts with dedicated delineations.