Detection Of Esophageal Cancers On Chest CT By Deep Learning

PurposeTo develop a deep learning-based model for helping radiologists find incidental esophageal cancers,thereby improving incidental detection of esophageal cancers in unenhanced chest CT examination.Material and MethodsWe retrospectively collected 141 patients with esophageal cancers(mean age: 57.4,range: 34-87)and 273 non-esophageal cancer subjects(mean age: 41.7,range: 18-73)in unenhanced chest CT images from January 2017 to March 2019.These Data were used to establish a convolutional neural(CNN)network model for diagnosing esophageal cancer.The CNN model is a V-Net segmentation network that can segment the esophagus and localize the thickening positions of esophageal lesions.To validate this model,another 52 missed cases of patients with esophageal cancers(mean age: 58.8,range: 23-85)and 48 normal esophagus(mean age: 42.1,range: 19-62)were collected to evaluate the performance of the deep learning based model and radiologists with or without assistance of deep learning respectively.ResultsThe sensitivity and specificity of the esophageal cancer detection model were88.8% and 90.9%,respectively.Of the 52 missed esophageal cancer cases and 48 normal cases,the sensitivity,specificity and accuracy of the deep learning esophageal cancer detection model were69%,61%,65%,respectively,and the radiologists’ independent reading results were sensitivity 25%,31%,27%,specificity 78%,75%,75%,and accuracy 53%,54%,53% in the absence of model assistance.With the aid of the model,the radiologists’ reading results were sensitivity 77%,81%,75%,specificity 75%,74%,74% and accuracy 76%,77%,75%respectively.ConclusionDeep learning-based model can effectively detect esophageal cancers in unenhanced chest CT scan to help improve incidental detection of esophageal cancer on CT.