Work

EndoPro: A Secure, Edge-Optimized Heuristic-Augmented Deep Learning System for Real-Time Gastrointestinal Anomaly Detection

Year: 2026
Type: conference-paper
Abstract: Although deep learning has demonstrated remarkable theoretical accuracy in gastrointestinal diagnostics, cloud-infrastructure dependencies, expensive GPU requirements, and neural networks’ vulnerability to Out-of-Distribution (OOD) environmental noise significantly impede real-world clinical deployment. EndoPro, a completely localised, offline-first edge AI system for real-time endoscopic analysis on commodity CPU hardware, is presented in this study. The intelligence layer was trained using an EfficientNetV2-Large architecture on a mathematically trimmed, 16-class subset of the Hyper-Kvasir dataset (10,478 pictures) employing sophisticated spatial augmentations in order to create a maximum-accuracy diagnostic baseline. This method produced an unseen test accuracy of 89.58% and a peak validation accuracy of 94.91%. The real-time deployment step uses edge-optimized cross-compilation using the Open Neural Network Exchange (ONNX) runtime in conjunction with a novel RGB-dominance heuristic pre-filter to address the significant computational limitations present in large deep learning designs. By successfully separating tensor processing from the primary application thread, this architectural technique achieves continuous inference speeds of more than 30 frames per second (FPS). EndoPro bridges the crucial gap between resource-constrained clinical viability and high-performance diagnostic AI by integrating hardware-level USB encrypted authentication for absolute data sovereignty.
Language: English
Open Access status: green
OpenAlex ID: W7166318667
FWCI:
Cites:
Cited by:
Related to: 10