Identifying cell types based on expression profiles is a pillar of single cell analysis. Existing machine-learning methods identify predictive features from annotated training data, which are often not available in early-stage studies. This can lead to overfitting and inferior performance when applied to new data. To address these challenges we present scROSHI, which utilizes previously obtained cell type-specific gene lists and does not require training or the existence of annotated data. By respecting the hierarchical nature of cell type relationships and assigning cells consecutively to more specialized identities, excellent prediction performance is achieved. In a benchmark based on publicly available PBMC data sets, scROSHI outperforms competing methods when training data are limited or the diversity between experiments is large.
Accurate risk prediction of distant recurrence (DR) is crucial for personalized adjuvant systemic therapy of endometrial cancer (EC) stage I-III patients, as DR is associated with a 5-year overall survival of 10-20%. Risk stratification and treatment recommendation are currently based on histopathological and molecular markers, which is challenging due to high inter-observer variability and testing costs respectively. Deep Learning (DL) models can predict prognosis by identifying relevant visual features from H&E whole slide images (WSIs) at different resolutions without prior assumptions. Here, we developed and tested the first interpretable state-of-the-art DL model for WSI-based risk prediction of DR of stage I-III EC (DeREC) from the randomized PORTEC-1/-2/-3 trials and three clinical cohorts with long-term follow-up data. We used one representative H&E WSI each from 1761 EC patients, excluding those who received adjuvant chemotherapy as it lowers the risk of DR. We randomly sampled 20% as a held-out internal test set (N=353 with 62 events; 8.45 year median follow-up) and performed a 5-fold cross-validation on the training set (N=1408). WSIs were partitioned into 360 micron patches at 40x magnification. DeREC combined self-supervised representation learning of patches using a multi-resolution vision transformer and a WSI-level graph attention-based time-to-event prediction model. The model performance of correctly ranking patients by predicted risk scores and true time to DR was measured with the concordance-index and compared with a Cox’ Proportional Hazards (CPH) model fitted on histopathological variables (histotype, grade, lymphovascular space invasion, stage). Discriminative quality of the predicted risk groups was investigated with Kaplan-Meier analysis and the log-rank test. Most predictive patches by predicted risk groups were reviewed by an expert gynecopathologist for identification of prognostic morphological features. DeREC achieved a concordance-index of 0.764 [95%CI 0.754-0.773] on 5-fold cross validation and 0.757 on the test set, as compared to 0.704 [95%CI 0.662-0.746] with CPH. Predicted risk groups around quartiles 1 and 3 accurately stratified patients between low (N=89), intermediate (N=175), high (N=89) risk of DR (p<0.0001). Among the predicted low-risk group only 3 (3.37%) patients relapsed whereas intermediate and high-risk groups counted 27 (15.43%) and 32 (35.96%) events respectively. DeREC is the first DL model accurately distinguishing EC patients at high risk of DR from those at low risk using one H&E WSI, which would aid decisions on adjuvant treatment. DeREC outperformed standard statistical prediction methods using histopathological variables, indicating that it identified prognostic visual features which can be further investigated. Future development includes the integration of clinicopathological and molecular information. Citation Format: Sarah Fremond, Sonali Andani, Jurriaan Barkey Wolf, Gitte Ørtoft, Estrid Høgdall, Jouke Dijkstra, Jan J. Jobsen, Ina M. Jürgenliemk-Schulz, Ludy CHW Lutgens, Melanie E. Powell, Naveena Singh, Linda R. Mileshkin, Helen J. Mackay, Alexandra Leary, Dionyssios Katsaros, Hans W. Nijman, Stephanie M. de Boer, Remi A. Nout, Vincent T.H.B.M Smit, Carien L. Creutzberg, Nanda Horeweg, Viktor H. Koelzer, Tjalling Bosse. Deep learning risk prediction model of distant recurrence from H&E endometrial cancer slides. [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2023; Part 1 (Regular and Invited Abstracts); 2023 Apr 14-19; Orlando, FL. Philadelphia (PA): AACR; Cancer Res 2023;83(7_Suppl):Abstract nr 5695.
scite is a Brooklyn-based organization that helps researchers better discover and understand research articles through Smart Citations–citations that display the context of the citation and describe whether the article provides supporting or contrasting evidence. scite is used by students and researchers from around the world and is funded in part by the National Science Foundation and the National Institute on Drug Abuse of the National Institutes of Health.
customersupport@researchsolutions.com
10624 S. Eastern Ave., Ste. A-614
Henderson, NV 89052, USA
This site is protected by reCAPTCHA and the Google Privacy Policy and Terms of Service apply.
Copyright © 2024 scite LLC. All rights reserved.
Made with 💙 for researchers
Part of the Research Solutions Family.