Efficient Synchronous Real-Time CADe for Multicategory Lesions in Gastroscopy by Using Multiclass Detection Model.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Gastroenterology

Background:

• Endoscopic examinations often require identifying multiple types of gastrointestinal lesions simultaneously.
• Existing computer-aided detection (CADe) systems may not efficiently handle multicategory lesion detection in real-time.

Purpose of the Study:

• To develop an efficient, synchronous, real-time CADe system for detecting multiple categories of gastrointestinal lesions.
• To compare a proposed multiclass YOLOv5 detection model against joint detection systems using single-class models.

Main Methods:

• A multiclass detection model based on YOLOv5 was developed for synchronous, real-time lesion detection.
• A retrospective dataset of 31,117 images from 3,747 patients was utilized.
• Online data augmentation and multiple loss functions were employed for model training.

Main Results:

• The proposed CADe system achieved 98% precision, 89% recall, and 90.2% mAP in detecting cancers, gastrointestinal stromal tumors, polyps, and ulcers.
• The system demonstrated a detection speed of 47 frames per second with 0.04s latency.
• The multiclass YOLOv5 system outperformed comparative joint detection systems in accuracy, speed, and latency.

Conclusions:

• The developed synchronous real-time CADe system with a multiclass detection model effectively detects multiple gastrointestinal lesions with high accuracy and efficiency.
• This system enhances the clinical application of CADe in endoscopy, offering a more efficient use of labeled medical images compared to multiple single-category models.