Beyond Binary: Four-Class Risk Stratification from Gastrointestinal Endoscopy Using Asymmetric-Cost Lightweight CNN–Transformer Learning

<title>Abstract</title> <p>Gastrointestinal cancers account for more than 3.5 million deaths annually, yet existing AI-assisted endoscopy systems reduce the clinical decision to a binary lesion-present/absent output misaligned with published clinical practice guidelines. We present a lightweight CNN–Transformer benchmark for four-class gastrointesti- nal lesion risk stratification aligned with American College of Gastroenterology (ACG) and European Society of Gastrointestinal Endoscopy (ESGE) guidelines. We mapped 21 of 23 HyperKvasir finding classes to four clinically actionable risk tiers—Normal, Inflammatory, Pre-malignant, and High-Risk—and trained three lightweight architectures (DenseNet-121, EfficientNet-B0, DeiT-Tiny; all &lt;8 M pa- rameters) under a novel Asymmetric Endoscopy Loss (AEL) that assigns a 5× misclassification penalty to High-Risk lesions. CLAHE preprocessing, Weight- edRandomSampler, and MC Dropout uncertainty quantification (T = 30 passes, τ = 0.75) complete the pipeline. DenseNet-121 achieved macro F1 = 0.84 on Hyper- Kvasir; EfficientNet-B0 leads zero-shot generalization at 0.78 on the independent Kvasir-v2 cohort. Zero High-Risk lesions were missed across all architectures un- der the MC Dropout referral protocol—a guarantee enforced by mandatory esca- lation of all Pre-malignant and High-Risk predictions—with 44.9 % of cases auto- cleared without endoscopist review. All models produced confidence estimates in the range ECE = 0.08–0.09. GradCAM saliency maps confirm that risk-tier decisions are grounded in clinically relevant mucosal features.</p>