S684 Real-Time Colorectal Polyp Detection and Stratification Using Combined Spatial Embeddings and Geometric Analysis

Introduction: Colorectal polyps represent critical precancerous lesions, and accurate detection with malignancy risk assessment during colonoscopy can markedly reduce cancer mortality. Most automated systems focus exclusively on visual appearance and lack reliable stratification of lesion risk. We aimed to develop a unified framework that integrates deep-learning–based appearance embeddings with geometric descriptors to detect polyps and assign malignancy risk in real time. Methods: We sourced 10 662 HyperKvasir endoscopic images labeled polyp or non-polyp. We resized frames to 224 × 224 pixels, normalized to ImageNet statistics, and applied random horizontal flips with brightness and contrast adjustments. We extracted 768-dimensional embeddings via a pretrained Vision Transformer (ViT-Base) and fed them into a feedforward network with 2 hidden layers of 512 and 256 units, ReLU activation, batch normalization, and 0.3 dropout. We trained with Adam (learning rate 1 × 10⁻⁴, weight decay 1 × 10⁻⁵) for 100 epochs (batch size 32) on an 80/10/10 stratified split. For risk stratification, we sampled 1 000 polyp instances, used bounding-box area and aspect-ratio thresholds to define low/moderate/high risk tiers, concatenated these features with embeddings (total 770 dimensions), and trained an identical network. We reported accuracy, precision, recall, F1-score, and receiver operating characteristic area under the curve (ROC AUC). Results: Polyp detection achieved 97.6% accuracy, 93.2% precision, 94.1% recall, F1-score 93.6%, and ROC AUC 0.995; misclassified frames represented fewer than 2% and showed no class bias. Risk stratification reached 96.5% accuracy, F1-score 0.96, and ROC AUC 0.999. Confusion-matrix analysis demonstrated clear separation of low- and high-risk tiers; moderate-risk cases accounted for <3% of errors. Fivefold cross-validation confirmed stability (standard deviation <1% for all metrics). Conclusion: The dual-feature framework delivered 97.6% detection accuracy and a 0.995 ROC AUC for polyp identification, alongside 96.5% accuracy and a 0.999 ROC AUC for risk stratification. Cross-validation showed metric stability (±<1%), and confusion-matrix analysis confirmed clear separation between risk tiers. Integrating spatial embeddings with geometric cues may enhance endoscopic decision support and reduce diagnostic variability.