Lightweight deep learning model for gastrointestinal precancerous lesion screening with attention enhancement.

Introduction: Gastrointestinal precancerous lesions, including colorectal polyps and gastric intraepithelial neoplasia, share highly similar morphological and pathological characteristics. In this study, we validate our lightweight attention-enhanced model on the Kvasir colorectal dataset, which serves as a representative benchmark for gastrointestinal lesion classification. Objective: Early detection of gastric precancerous lesions is crucial for reducing gastric cancer mortality. This study aimed to address the trade-off between lightweight performance, classification accuracy and interpretability of existing AI models for endoscopic image analysis, and develop a practical attention-enhanced lightweight model for gastrointestinal precancerous lesion classification to support clinical application. Methods: An attention-enhanced lightweight MobileNetV3 model was proposed by integrating a Spatial-Channel Attention (SCA) module after the last convolutional layer of the backbone for dual spatial and channel weighting to enhance lesion-focused feature extraction. Ablation experiments were conducted to verify the SCA module's effectiveness by comparing models with and without the module. GradCAM was used to visualize the model's decision-making process. The model was evaluated on the Kvasir dataset containing 3,000 endoscopic images (3 classes, normal gastrointestinal mucosa, polyp, ulcerative colitis) with a 7:2:1 train-validation-test split. The training adopted Adam optimizer (lr = 1e-4), StepLR scheduler and early stopping (patience = 5) to ensure stability. Results: The proposed model achieved an overall accuracy of 99.10% and a weighted F1-score of 99.10%. For polyp (the key high-risk precancerous lesion), it reached 100% precision and 98% recall. Ablation experiments confirmed that the SCA module stably improved core metrics by 0.07% without increasing computational redundancy. The model maintained ultra-lightweight characteristics with only 1.05 M total parameters and 3.99 MB parameter storage, supporting real-time deployment on endoscopic devices. GradCAM heatmaps showed that the SCA module enabled the model to precisely focus on clinically relevant regions (e.g., polyp edges), consistent with endoscopists' observation habits. Conclusion: The proposed model effectively balances lightweight performance, high classification accuracy and interpretability. Its excellent polyp recognition performance reduces the risks of misdiagnosis and missed diagnosis in clinical screening, while the lightweight design and GradCAM-based interpretability break the key barriers to the clinical adoption of AI in gastric cancer early diagnosis, providing a feasible AI-assisted tool for gastric precancerous lesion screening.