Artificial intelligence-assisted endoscopy in the detection of early gastrointestinal cancer: Progress, challenges, and future directions

Table 1 Explanation of deep convolutional neural network architectures

Architecture name	Core innovation/structural features	Relevance in medical field
VGGNet	Adopts a concise structure of “stacked small convolutional kernels (3 × 3) + pooling layers”, enhancing feature extraction capability by increasing network depth	A classic model for basic feature extraction in medical images, suitable for preliminary lesion detection and medical image classification (e.g., X-ray disease screening), laying the foundation for subsequent architectures in medical AI
ResNet	Introduces “residual connections” (cross-layer feature transmission) to solve the gradient vanishing problem in deep network training, enabling the construction of ultra-deep networks	Significantly improves feature extraction accuracy for complex medical images, applicable to pathological section analysis and 3D medical image segmentation (e.g., tumor boundary extraction), serving as a core architecture for disease diagnosis models
DenseNet	Employs “dense connections” (direct feature sharing across all layers) to enhance feature propagation efficiency and reduce parameter redundancy	Excels in fine-grained analysis of medical images, such as micro-lesion recognition and multi-modal medical image fusion (e.g., combining CT and MRI images), demonstrating distinct advantages in precision medical diagnosis

VGGNet: Visual geometry group network; ResNet: Residual network; DenseNet: Dense convolutional network; AI: Artificial intelligence; 3D: Three dimensional; CT: Computed tomography; MRI: Magnetic resonance imaging.

Table 2 Summary of comparisons of major artificial intelligence-assisted endoscopy systems

System name	Target site and function	Key performance metrics	Validation status and characteristics
Deep learning-based endoscopy systems	Esophagus, stomach: Early cancer detection and diagnosis	Sensitivity for early gastric cancer > 90%, specificity > 80%	Mostly in clinical research phase: Validation often involves single-center or retrospective studies; demonstrates potential to match or surpass human experts in specific tasks
		Detection accuracy for early esophageal cancer comparable to expert endoscopists
AI-assisted capsule endoscopy systems	Small bowel: Automatic detection of ulcers, bleeding, polyps, etc.	Sensitivity for small bowel lesions > 95%, specificity > 90%	Validated by multicenter prospective studies: Some systems have received regulatory approval and are in clinical use; aims to address the inefficiency of analyzing large CE image volumes
		Significantly increases reading speed, reducing physician workload by > 70%
CADe colonoscopy systems	Colorectum: Real-time polyp detection (CADe)	Increases adenoma detection rate by an absolute 5%-10%	Some systems approved by FDA, CE, NMPA: Supported by the highest level of evidence (multicenter RCTs); integrated into commercial endoscopy platforms; value is pronounced in community practice settings
		Particularly effective for detecting small polyps (< 5 mm) and flat adenomas
CADx colonoscopy systems	Colorectum: Real-time polyp characterization (CADx)	Accuracy for optical diagnosis of adenomatous polyps > 90%	Some features approved and commercialized: Integrated with CADe systems; aims to provide “see-and-diagnose” capability, reducing unnecessary polypectomies and screening costs
		Enables reliable “diagnose-and-leave” or “resect-and-discard” strategies with > 90% confidence
AI-assisted laryngoscopy/pharyngeal diagnosis systems	Pharynx, larynx: Early cancer detection	Sensitivity for laryngopharyngeal cancer 90%-93%, specificity > 92%	Primarily in prospective research or pilot project phase: Sample sizes are relatively small, but shows great promise for multimodal AI in complex anatomical sites
		Capable of multimodal analysis integrating voice signals and images
AI-assisted high-resolution anoscopy	Anal canal: Detection of HSIL	Shows high accuracy in differentiating HSIL from other conditions	Research is very preliminary and exploratory: Limited sample sizes; a promising tool for screening specific high-risk populations but requires further validation
		Can integrate HPV status for risk prediction

AI: Artificial intelligence; CE: Contrast-enhanced; CADe: Computer-aided detection; CADx: Computer-aided diagnosis; HSIL: High-grade squamous intraepithelial lesion; HPV: Human papillomavirus; FDA: Food and Drug Administration; NMPA: National Medical Products Administration; RCTs: Randomized controlled trials.