Deep learning Instance Segmentation on Esophageal Squamous Cell Carcinoma detection

ObjectiveEsophageal cancer (EC) is a prevalent malignancy worldwide. Early-stage esophageal cancer (EEC) diagnostics is crucial for improving patient survival. However, EC is highly aggressive with a poor prognosis, even for experienced endoscopists. To address these problems, this study aims to develop a novel computer-aided diagnosis (CAD) method to improve the accuracy and efficiency of EEC diagnostics. MethodsThree-way fusion CAD method that employs multiple frameworks, including the hybrid task cascade ResNeXt101 with deformable convolutional networks, to accurately detect EC. Our method incorporates dual annotation categories on ME-NBI imaging from a local perspective and one category on LCE imaging from an broader perspective. This integration provides a substantial improvement of accuracy over traditional CAD technologies. ResultsOur three-way fusion CAD method achieved top performances of 0.923 mAP on ME-NBI and 0.862 mAP on LCE, demonstrating superior diagnostic performance compared to traditional CAD methods. Furthermore, the treatment boundary mAP is expected to be even higher by definition in clinical settings. Our method also achieved promising precision and recall rates of 93.98% and 93.05% for ME-NBI, and 82.89% and 88.32% for LCE, respectively. ConclusionsOur novel three-way fusion CAD method accurately detects EC in both ME-NBI and LCE imaging, providing accurate treatment boundaries on both image and patient levels. Our approach shows potential for clinical application, with promising mAP, precision, and recall rates. Further work will focus on collecting and analyzing patient data to improve the methods real-time performance in clinical settings.