Early esophageal adenocarcinoma detection using deep learning methods.

Area of Science:

• Medical imaging analysis
• Artificial intelligence in healthcare
• Gastroenterology

Background:

• Esophageal adenocarcinoma (EAC) diagnosis relies on endoscopic visualization.
• Accurate identification of abnormal regions is crucial for timely treatment.
• Manual review of endoscopic images can be time-consuming and subjective.

Purpose of the Study:

• To adapt and evaluate deep learning object detection models for automatic EAC region identification.
• To compare the performance of various state-of-the-art object detection methods.
• To assess the utility of these methods in high-definition white light endoscopy (HD-WLE) images.

Main Methods:

• Utilized Convolutional Neural Networks (CNNs) with VGG'16 backbone.
• Adapted and tested Regional-based Convolutional Neural Network (R-CNN), Fast R-CNN, Faster R-CNN, and Single-Shot Multibox Detector (SSD).
• Evaluated models on 100 manually annotated HD-WLE images from 39 patients.

Main Results:

• Single-Shot Multibox Detector (SSD) and Faster R-CNN demonstrated strong performance.
• SSD achieved the highest performance with 0.96 sensitivity, 0.92 specificity, and 0.94 F-measure.
• Faster R-CNN achieved an Average Recall Rate of 0.83 for accurate EAC region localization.

Conclusions:

• Deep learning object detection methods can effectively locate esophageal abnormalities in endoscopic images.
• Automatic detection is a vital step for early EAC detection and treatment monitoring.
• These AI tools can potentially enhance tumor segmentation and outcome assessment.