Development of a deep learning-based automatic detection model for gastroesophageal varices using transient elastography

Background: and Purpose: This study investigates the application of deep learning in the early detection and risk assessment of gastroesophageal varices (GOV), focusing on non-invasive methods using transient elastography. Methods: The study employs Dense Net 201 and an enhanced Dense Net 201-with-SSMV (Spleen Stiffness Measurement Value) models. These models are trained and tested on a dataset comprising clinical and transient elastography data from patients. The methodology includes preprocessing of data, model training, and validation. To compare the performance of these models, metrics such as accuracy, sensitivity, specificity, and area under the ROC curve are utilized. Additionally, a thorough cross-validation process is implemented to ensure the robustness and generalizability of the models. The study aims to establish the superiority of the Dense Net 201-with-SSMV model over the standard Dense Net 201 by demonstrating enhanced predictive performance in detecting gastroesophageal varices. Results: The Dense Net 201-with-SSMV model demonstrated superior performance with a training accuracy of 94.25%, sensitivity of 91.70%, specificity of 96.19%, and AUC of 0.988 (95% CI: 0.982-0.994). In testing, it achieved an accuracy of 90.15%, sensitivity of 84.03%, specificity of 94.76%, and AUC of 0.946 (95% CI: 0.919-0.973), outperforming the Dense Net 201 model. The study demonstrates the superior performance of the Dense Net 201-with-SSMV model over the standard Dense Net 201, particularly in accuracy and predictive capabilities. Conclusion: The integration of clinical indicators with deep learning models offers a promising, non-invasive approach for GOV detection, with potential implications for liver disease management. The study also identifies the need for larger datasets and exploration of additional clinical indicators to enhance model performance