Evaluating reproducibility of AI algorithms in digital pathology with DAPPER

Artificial Intelligence is exponentially increasing its impact on healthcare. As deep learning is mastering computer vision tasks, its application to digital pathology is natural, with the promise of aiding in routine reporting and standardizing results across trials. Deep learning features inferred from digital pathology scans can improve validity and robustness of current clinico-pathological features, up to identifying novel histological patterns, e.g., from tumor infiltrating lymphocytes. In this study, we examine the issue of evaluating accuracy of predictive models from deep learning features in digital pathology, as an hallmark of reproducibility. We introduce the DAPPER framework for validation based on a rigorous Data Analysis Plan derived from the FDA's MAQC project, designed to analyze causes of variability in predictive biomarkers. We apply the framework on models that identify tissue of origin on 787 Whole Slide Images from the Genotype-Tissue Expression (GTEx) project. We test three different deep learning architectures (VGG, ResNet, Inception) as feature extractors and three classifiers (a fully connected multilayer, Support Vector Machine and Random Forests) and work with four datasets (5, 10, 20 or 30 classes), for a total of 53, 000 tiles at 512 x 512 resolution. We analyze accuracy and feature stability of the machine learning classifiers, also demonstrating the need for diagnostic tests (e.g., random labels) to identify selection bias and risks for reproducibility. Further, we use the deep features from the VGG model from GTEx on the KIMIA24 dataset for identification of slide of origin (24 classes) to train a classifier on 1, 060 annotated tiles and validated on 265 unseen ones. The DAPPER software, including its deep learning pipeline and the Histological ImagingNewsy Tiles (HINT) benchmark dataset derived from GTEx, is released as a basis for standardization and validation initiatives in AI for digital pathology. Author summary In this study, we examine the issue of evaluating accuracy of predictive models from deep learning features in digital pathology, as an hallmark of reproducibility. It is indeed a top priority that reproducibility-by-design gets adopted as standard practice in building and validating AI methods in the healthcare domain. Here we introduce DAPPER, a first framework to evaluate deep features and classifiers in digital pathology, based on a rigorous data analysis plan originally developed in the FDA's MAQC initiative for predictive biomarkers from massive omics data. We apply DAPPER on models trained to identify tissue of origin from the HINT benchmark dataset of 53, 000 tiles from 787 Whole Slide Images in the Genotype-Tissue Expression (GTEx) project, available at the web address https://gtexportal.org. We analyze accuracy and feature stability of different deep learning architectures (VGG, ResNet and Inception) as feature extractors and classifiers (a fully connected multilayer, Support Vector Machine and Random Forests) on up to 20 classes. Further, we use the deep features from the VGG model (trained on HINT) on the 1, 300 annotated tiles of the KIMIA24 dataset for identification of slide of origin (24 classes). The DAPPER software is available together with the scripts to generate the HINT benchmark dataset.