Generating synthetic data in digital pathology through diffusion models: a multifaceted approach to evaluation

Abstract

Synthetic data has recently risen as a new precious item in the computational pathologist’s toolbox, supporting several tasks such as helping with data scarcity or augmenting training set in deep learning. Nonetheless, the use of such novel resources requires a carefully planned construction and evaluation, to avoid pitfalls such as the generation of clinically meaningless artifacts.As the major outcome described in the current manuscript, a novel full stack pipeline is introduced for the generation and evaluation of synthetic pathology data powered by a diffusion model. The workflow features, as characterizing elements, a new multifaceted evaluation strategy with an embedded explainability procedure effectively tackling two critical aspects of the use of synthetic data in health-related domains.An ensemble-like strategy is adopted for the evaluation of the produced data, with the threefold aim of assessing the similarity of real and synthetic data through a set of well-established metrics, evaluating the practical usability of the generated images in deep learning models complemented by explainable AI methods, and validating their histopathological realism through a dedicated questionnaire answered by three professional pathologists.The pipeline is demonstrated on the public GTEx dataset of 650 WSIs, including five different tissues, conditioning the training step of the underlying diffusion model. An equal number of tiles from each of these five tissues are then generated. Finally, the reliability of the generated data is assessed using the proposed evaluation pipeline, with encouraging results. We show that each of these evaluation steps are necessary as they provide complementary information on the generated data’s quality.Overall, all the aforementioned features characterize the proposed workflow as a fully-fledged solution for generative AI in digital pathology representing a potentially useful tool for the digital pathology community in their transition towards digitalization and data-driven modeling.