Forecasting the “T” Stage of Esophageal Cancer by Deep Learning Methods: A Pilot Study

Abstract

Research motivation: Staging esophageal cancer is of paramount importance for treatment. With conventional methods, accuracy of staging is low, we aimed to improve the accuracy of the “T” stage of esophageal cancer by using deep learning techniques. Method/Material: Clinically diagnosed esophageal cancer patients were prospectively observed and their data were collected. jpeg images were collected from the Computed Tomography of patients. 80% of the data were used for training and 20% for tests. Pathology results were used as the gold standard in the training of deep learning algorithms. EfficientNetB7 and ResNet152V2 models were used in the study. Both architectures with convolutional neural networks have Convolutional layers, pool layers, and fully connected layers. Results: A total of 477 images of 50 patients were analyzed. EfficientNetB7 makes predictions with a total of 64,107,931 parameters, and ResNet152V2 58,339,844 parameters within seconds (2[Formula: see text]s) at rates close to the accuracy offered by humans. With the EfficientNetB7 architecture, one of the Convolutional Neural Networks used in this study, 90% accuracy was achieved in the “T” staging of esophageal cancer. Conclusion: Despite the very limited dataset, deep learning algorithms can perform effective and reliable staging under the supervision of an experienced radiologist. With more datasets, the precision of the estimation can increase.