Built to last? Reproducibility and Reusability of Deep Learning Algorithms in Computational Pathology

Abstract

Recent progress in computational pathology has been driven by deep learning. While code and data availability are essential to reproduce findings from preceding publications, ensuring a deep learning model’s reusability is more challenging. For that, the codebase should be well-documented and easy to integrate in existing workflows, and models should be robust towards noise and generalizable towards data from different sources. Strikingly, only a few computational pathology algorithms have been reused by other researchers so far, let alone employed in a clinical setting.To assess the current state of reproducibility and reusability of computational pathology algorithms, we evaluated peer-reviewed articles available in Pubmed, published between January 2019 and March 2021, in five use cases: stain normalization, tissue type segmentation, evaluation of cell-level features, genetic alteration prediction, and direct extraction of grading, staging, and prognostic information. We compiled criteria for data and code availability, and for statistical result analysis and assessed them in 161 publications. We found that only one quarter (42 out of 161 publications) made code publicly available and thus fulfilled our minimum requirement for reproducibility and reusability. Among these 42 papers, three quarters (30 out of 42) analyzed their results statistically, less than half (20 out of 42) have released their trained model weights, and only about a third (16 out of 42) used an independent cohort for evaluation.This review highlights candidates for reproducible and reusable algorithms in computational pathology. It is intended for both pathologists interested in deep learning, and researchers applying deep learning algorithms to computational pathology challenges. We provide a list of reusable data handling tools and a detailed overview of the publications together with our criteria for reproducibility and reusability.

Closing the gap in the clinical adoption of computational pathology: a standardized, open-source framework to integrate deep-learning algorithms into the laboratory information system

“…5,10 Many commercial DL algorithms run on cloud servers, and often rely on proprietary viewers, whose integration within existing AP-LIS is non-trivial. Most of the publicly available DL models on the other hand are not reusable, 11,12 not even in fully-digitized diagnostic workflows, nor accessible to pathologists without programming skills. 13 Even for the minority of DL algorithms that have been published in a format suitable for use by other researchers, their integration into the AP-LIS remains challenging.…”