Abstract / Summary
Foundation models are trained on massive amounts of data to capture complex patterns in images. Subsequently, a wide range of downstream tasks can be adopted with minimal computational resources. We have developed HistoEncoder, a foundation model for prostate cancer digital pathology by pre-training on 48 million prostate tissue tile images. HistoEncoder allows automated extraction of histological features highly predictive of Gleason patterns achieving comparable performance with substantially larger pan-cancer foundation models while being much more efficient. By fine-tuning the model with a small amount of data and computational resources, we describe two clinical use cases for HistoEncoder. First, HistoEncoder can be used to automatically annotate large-scale datasets with high accuracy. Second, we show that HistoEncoder-derived histology clusters contain prognostic information of a similar magnitude to Gleason grading in internal cross-validation. Lightweight foundation models such as HistoEncoder allow organizations to build effective clinical software tools without the need for extensive datasets and heavy computing.