Search

Objective: Previous reports of deep learning-assisted assessment of Mayo endoscopic subscore (MES) in ulcerative colitis have only explored the ability to distinguish disease remission (MES 0/1) from severe disease (MES 2/3) or inactive disease (MES 0) from active disease (MES 1-3). We sought to explore the utility of deep learning models in the automated grading of each individual MES in ulcerative colitis., Methods: In this retrospective study, a total of 777 representative still images of endoscopies from 777 patients with clinically active ulcerative colitis were graded using the MES by two physicians. Each image was assigned an MES of 1, 2, or 3. A 101-layer convolutional neural network model was trained and validated on 90% of the data, while 10% was left for a holdout test set. Model discrimination was assessed by calculating the area under the curve (AUC) of a receiver operating characteristic as well as standard measures of accuracy, specificity, sensitivity, positive predictive value (PPV), and negative predictive value (NPV)., Results: In the holdout test set, the final model classified MES 3 disease with an AUC of 0.96, MES 2 disease with an AUC of 0.86, and MES 1 disease with an AUC 0.89. Overall accuracy was 77.2%. Across MES 1, 2, and 3, average specificity was 85.7%, average sensitivity was 72.4%, average PPV was 77.7%, and the average NPV was 87.0%., Conclusion: We have demonstrated a deep learning model was able to robustly classify individual grades of endoscopic disease severity among patients with ulcerative colitis., (Copyright © 2020 Wolters Kluwer Health, Inc. All rights reserved.)

Background: Genitourinary rhabdomyosarcoma (GU-RMS) is a rare, pediatric malignancy originating from embryonic mesenchyme. Current approaches to prognostication rely upon conventional statistical methods such as Cox proportional hazards (CPH) models and have suboptimal predictive ability. Given the success of deep learning approaches in other specialties, we sought to develop and compare deep learning models with CPH models for the prediction of 5-year survival in pediatric GU-RMS patients., Methods: Patients less than 20 years of age with GU-RMS were identified within the Surveillance, Epidemiology, and End Results (SEER) database (1998-2011). Deep neural networks (DNN) were trained and tested on an 80/20 split of the dataset in a 5-fold cross-validated fashion. Multivariable CPH models were developed in parallel. The primary outcomes were 5-year overall survival (OS) and disease-specific survival (DSS). Variables used for prediction were age, sex, race, primary site, histology, degree of tumor extension, tumor size, receipt of surgery, and receipt of radiation. Receiver operating characteristic curve analysis was conducted, and DNN models were tested for calibration., Results: 277 patients were included. The area under the curve (AUC) for the DNN models was 0.93 for OS and 0.91 for DSS. AUC for the CPH models was 0.82 for OS and 0.84 for DSS. The DNN models were well-calibrated: OS model (slope = 1.02, intercept = -0.06) and DSS model (slope = 0.79, intercept = 0.21)., Conclusions: A deep learning-based model demonstrated excellent performance, superior to that of CPH models, in the prediction of pediatric GU-RMS survival. Deep learning approaches may enable improved prognostication for patients with rare cancers., (Copyright © 2020 Elsevier Ltd. All rights reserved.)