Lightweight deep learning model for gastrointestinal precancerous lesion screening with attention enhancement
-
By
-
Shuai Chen
-
Jingyao Cai
-
Zhixiang Wu
-
Xiangyu Liu
-
Qing Wang
-
Liming Zhou
-
June 4, 2026
-
Clinical Scorecard: Efficient Deep Learning Framework with Attention Mechanism for Screening Gastrointestinal Precancerous Lesions
At a Glance
| Category | Detail |
|---|
| Condition | Gastrointestinal precancerous lesions |
| Key Mechanisms | Attention-enhanced lightweight MobileNetV3 model with Spatial-Channel Attention (SCA) module |
| Target Population | Patients at risk for gastric cancer due to precancerous lesions |
| Care Setting | Endoscopic screening and diagnosis |
Key Highlights
- Achieved 99.10% overall accuracy and 100% precision for polyp detection
- Utilized GradCAM for model interpretability and decision-making visualization
- Maintained ultra-lightweight characteristics with only 1.05 M parameters
- Improved core metrics by 0.07% with the SCA module without increasing computational load
- Supports real-time deployment on endoscopic devices
Guideline-Based Recommendations
Diagnosis
- Endoscopy is the gold standard for identifying gastric precancerous lesions.
Management
- Early detection and intervention can reduce the risk of malignant transformation by up to 80%.
Monitoring & Follow-up
- Regular screening for high-risk populations is essential.
Risks
- Misdiagnosis and missed diagnosis in clinical screening can occur without AI assistance.
Patient & Prescribing Data
Individuals undergoing endoscopic screening for gastric cancer risk
AI-assisted tools can enhance diagnostic accuracy and reduce clinician workload.
Clinical Best Practices
- Incorporate AI models with interpretability tools like GradCAM in clinical workflows.
- Utilize lightweight models to improve accessibility in resource-limited settings.
- Focus on enhancing feature discrimination between lesions and healthy tissue.
Related Resources & Content
