Direct prediction of genetic aberrations from pathology images in gastric cancer with swarm learning

Abstract

Background: Computational pathology uses deep learning (DL) to extract biomarkers from routine pathology slides. Large multicentric datasets improve performance, but such datasets are scarce for gastric cancer. This limitation could be overcome by Swarm Learning (SL).

Methods: Here, we report the results of a multicentric retrospective study of SL for prediction of molecular biomarkers in gastric cancer. We collected tissue samples with known microsatellite instability (MSI) and Epstein-Barr Virus (EBV) status from four patient cohorts from Switzerland, Germany, the UK and the USA, storing each dataset on a physically separate computer.

Results: On an external validation cohort, the SL-based classifier reached an area under the receiver operating curve (AUROC) of 0.8092 (± 0.0132) for MSI prediction and 0.8372 (± 0.0179) for EBV prediction. The centralized model, which was trained on all datasets on a single computer, reached a similar performance.

Conclusions: Our findings demonstrate the feasibility of SL-based molecular biomarkers in gastric cancer. In the future, SL could be used for collaborative training and, thus, improve the performance of these biomarkers. This may ultimately result in clinical-grade performance and generalizability.

Keywords: Artificial intelligence; Biomarker; Blockchain; Gastric cancer; Pathology; Swarm learning.

Conflict of interest statement

JNK declares consulting services for Owkin, France and Panakeia, UK. No other potential conflicts of interest are reported by any of the authors.

© 2022. The Author(s).

Figures

Fig. 1

Outline of this study. A Technical setup of the swarm learning experiment. B Distribution of training and testing set for the three experiments local models (each dataset is used to independently train a model), central models (all datasets are merged), and swarm model (all datasets are used to co-train a model without merging any raw data)

Fig. 2

MSI status prediction from pathology images in gastric cancer with swarm learning. A Classification performance (area under the receiver operating curve, AUROC) for prediction of MSI status on a patient level in the TCGA cohort. The results of three replicates per experiment are shown as a box plot. The box shows the median and quartiles as the whiskers expand to the rest of the distribution, with the exception of points identified as outliers. B Highly predictive image tiles for the Swarm Learning model for MSI and MSS, obtained from the first of three experiments. C Whole-slide prediction heatmaps for MSI and MSS in six patients. Abbreviations: w-chkpt weighted checkpoint of the swarm (= final swarm learning model), MSI microsatellite instable, MSS microsatellite stable

Fig. 3

EBV status prediction from pathology images in gastric cancer with swarm learning. A Classification performance (area under the receiver operating curve, AUROC) for prediction of EBV status on a patient level in the TCGA cohort. The results of three replicates per experiment are shown as a box plot, obtained from the first of three experiments. The box shows the median and quartiles as the whiskers expand to the rest of the distribution, with the exception of points identified as outliers. B Highly predictive image tiles for the Swarm Learning model for MSI and MSS. C Whole-slide prediction heatmaps for EBV positivity and negativity in six patients. Abbreviations: w-chkpt weighted checkpoint of the swarm (= final swarm learning model), EBV Epstein–Barr Virus, Pos. positive, Neg. negative