Introduction The aim of the present study was to develop deep learning-based models to assist in predicting the overall survival (OS) of patients with cervical cancer (CC) by directly analysing scanned conventional haematoxylin and eosin (H&E)-stained whole-slide images (WSIs).

Methods In total, 1161 HE-stained WSIs of primary cervical tumors from 405 patients who underwent radical CC surgery at the Fudan University Shanghai Cancer Center (FUSCC) between 2008 and 2014 were used in this retrospective study. The primary outcome was OS. Our deep learning model (CCOSNet) were developed using artificial intelligence (AI) for predicting outcome of CC. Performance was primarily evaluated using the sensitivity, specificity, and area under the receiver operating characteristic curve (AUROC).

Results We constructed and trained a multi-instance deep convolutional neural network (CCOSNet) based on a multiscale attention mechanism, in which an internal independent test set (50 patients in total) were used to evaluate the predictive performance of the network. Our network achieved an AUROC=0.88 in the cross-validation set and an AUROC=0.79 in the internal independent test set of the FUSCC cohort. The biomarker provided a hazard ratio for poor versus good prognosis of 7.058 (P = 0·009) in the primary analysis of the validation cohort.

Conclusion/Implications A clinically useful prognostic marker was developed using deep learning allied to digital scanning of conventional H&E-stained tumour tissue sections, which will offer assistance to choose appropriate treatment to improve the survival status of CC patients.