Predicting lymph node status based on deep learning on histopathological images from primary tumor of early-stage cervical squamous cell carcinoma (080)

Abstract

Objectives: To develop a new digital biomarker based on the convolutional neural network (CNN) analysis of primary tumor tissue, hence predicting lymph nodes (LNs) status in patients with early-stage cervical squamous cell carcinoma (CSCC). Methods: A total of 406 CSCC patients who underwent radical surgery between 2006 and 2014 were randomly selected, with their LNs status and hematoxylin-eosin (HE) stained primary tumor slides reviewed and confirmed by pathologists. Regions of interest (ROIs) were delineated by our pathology team. Slides were randomly divided into training, validation, and testing sets. With all Images preprocessed into patches, a CNN was trained to classify patches as LNs positive or negative. The CNN was further tested on a third independent population of 106 patients from The Cancer Genome Atlas (TCGA) database. Results: The cases included in the training set, validation set, and test set were 286, 60, and 60, respectively. Our classifier achieved an AUC of 0.86 (95% CI: 0.82-0.89) in cross-validation of slide-level LN status, 0.81 (95% CI: 0.77-0.86) on the test set, and 0.75 (95% CI: 0.71-0.82) on TCGA test set. For patients with positive LN, the digital signatures demonstrated good discrimination between patients with only pelvic LNs positive and patients with both pelvic and para-aortic LNs positive. The best performance model achieved an AUC of 0.91 (95% CI: 0.86-0.95) in the validation cohort and 0.87 (95% CI: 0.81-0.92) in the test cohort. Objectives: To develop a new digital biomarker based on the convolutional neural network (CNN) analysis of primary tumor tissue, hence predicting lymph nodes (LNs) status in patients with early-stage cervical squamous cell carcinoma (CSCC). Methods: A total of 406 CSCC patients who underwent radical surgery between 2006 and 2014 were randomly selected, with their LNs status and hematoxylin-eosin (HE) stained primary tumor slides reviewed and confirmed by pathologists. Regions of interest (ROIs) were delineated by our pathology team. Slides were randomly divided into training, validation, and testing sets. With all Images preprocessed into patches, a CNN was trained to classify patches as LNs positive or negative. The CNN was further tested on a third independent population of 106 patients from The Cancer Genome Atlas (TCGA) database. Results: The cases included in the training set, validation set, and test set were 286, 60, and 60, respectively. Our classifier achieved an AUC of 0.86 (95% CI: 0.82-0.89) in cross-validation of slide-level LN status, 0.81 (95% CI: 0.77-0.86) on the test set, and 0.75 (95% CI: 0.71-0.82) on TCGA test set. For patients with positive LN, the digital signatures demonstrated good discrimination between patients with only pelvic LNs positive and patients with both pelvic and para-aortic LNs positive. The best performance model achieved an AUC of 0.91 (95% CI: 0.86-0.95) in the validation cohort and 0.87 (95% CI: 0.81-0.92) in the test cohort.