Preparing Well for Esophageal Endoscopic Detection Using a Hybrid Model and Transfer Learning

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• Early detection of esophageal cancer via endoscopy is crucial but challenging due to ambiguous visual features and reliance on physician expertise.
• Current AI diagnostic models often require extensive datasets, limiting their use in low-data scenarios.
• Accurate differentiation from inflammatory esophageal conditions is critical for effective diagnosis.

Purpose of the Study:

• To develop and evaluate a novel AI-based computer-aided diagnosis system for esophageal cancer detection using endoscopic images.
• To address the challenge of limited data availability in medical imaging by employing transfer learning strategies.
• To enhance prediction accuracy through a hybrid model integrating EfficientNet and Vision Transformer networks.

Main Methods:

• Utilized transfer learning for data training strategies optimized for limited datasets.
• Developed a hybrid AI model combining EfficientNet and Vision Transformer architectures.
• Evaluated the model on a curated dataset of 1002 endoscopic images (650 white-light, 352 narrow-band).

Main Results:

• The hybrid AI model achieved high performance metrics: 96.32% accuracy, 96.44% precision, 95.70% recall, and 96.04% F1-score.
• Outperformed state-of-the-art models and individual network components in classification accuracy.
• Demonstrated superior prediction accuracy, compact model size, and adaptability to low-data environments.

Conclusions:

• The proposed AI platform offers a significant advancement in computer-aided endoscopic imaging for esophageal cancer diagnosis.
• The hybrid model shows strong potential for precise medical image classification in resource-limited settings.
• This research paves the way for improved early detection and treatment of esophageal cancer.