LoID-EEC: Localizing and Identifying Early Esophageal Cancer Based on Deep Learning in Screening Chromoendoscopy.

Esophageal cancer is one of the most common malignant tumors which responses for about 400,000 deaths each year. Early identifying lesions is critical for reducing esophageal cancer mortality and the overall esophageal cancer burden. However, identification of early esophageal cancerous lesions can be very challenging for clinicians owing to the mild clinical symptoms and lack of specificity of esophageal cancer. Consequently, precancer or subtle early neoplastic changes may not be evident, limiting the diagnostic accuracy. As a clinical assistance for early esophageal cancer identification, a deep learning framework referred to as the M-Deeplab model was proposed for the localization and recognition of esophageal mucosa lesion. The proposed M-Deeplab model was extended from the Deeplabv3+ model by employing an encoder-decoder structure for accuracy improvement. It achieves high-precision semantic segmentation for different staining degrees and different sizes of endoscopic images. The overall accuracy reaches 97.31% and the MIoU reaches 92.09%. Moreover, it takes only 0.05s to judge one image by the M-Deeplab model. The M-Deeplab model exhibits good performance both in accuracy and speed for early esophageal cancerous lesions identification, comparable to the experienced clinicians. As an assistance for the clinicians, the proposed model could possibly increase the early esophageal cancer diagnosis accuracy and decrease the misdiagnosis.