Reducing blind spots in esophagogastroduodenoscopy examinations using a novel deep learning model

Guangquan Wan,Guanghui Lian,Lan Yao

DOI: https://doi.org/10.1007/s00530-024-01259-2

IF: 3.9

2024-03-05

Multimedia Systems

Abstract:The intricate architecture of gastric anatomy coupled with the complexities inherent in esophagogastroduodenoscopy (EGD) procedures can lead to blind spots during examinations. These blind spots refer to anatomical locations not visualized during EGD examinations, potentially impacting timely diagnoses and treatments, and exacerbating patient conditions. Therefore, developing artificial intelligence (AI) for monitoring and reducing blind spots in EGD examinations is crucial. This study introduces MMCNet, a novel deep-learning model for classifying anatomical locations in EGD images. The model-based AI system can promptly alert the physician when it fails to recognize all anatomical locations in real-time EGD examinations, thereby enabling the monitoring and reduction of blind spots. To validate its efficacy, comprehensive experimental assessments compared MMCNet with established deep learning models. The results confirm MMCNet's high accuracy rate of 97.25% in recognizing anatomical locations in EGD images. Moreover, its notably compact memory size of 4.16M contributes to reduced memory requirements. With its accuracy and small model size, the model demonstrates significant potential as an effective tool for computer-assisted blind spot detection. Additionally, this study presents a comprehensive workflow for applying deep learning models to address practical issues, which can be easily adapted for similar tasks.

computer science, information systems, theory & methods