Next-Generation AI-Driven Early Detection of Gastric Ulcers Using High-Resolution Imaging, Deep Learning, and Multi-Scale Radiomic Feature Analysis

https://doi.org/10.5281/zenodo.18417526

Abstract

Gastric ulcers represent a major gastrointestinal disorder with substantial clinical burden, particularly when diagnosis is delayed or lesions are misinterpreted during early stages. Conventional diagnostic approaches rely heavily on endoscopic examination and expert visual assessment, which are inherently subjective, time-consuming, and susceptible to inter-observer variability. Recent advances in artificial intelligence (AI) and medical imaging offer promising opportunities to enhance diagnostic accuracy and enable early, objective, and scalable ulcer detection. In this study, we propose a next-generation AI-driven framework for the early detection of gastric ulcers by synergistically integrating high-resolution endoscopic imaging, deep learning–based feature learning, and multi-scale radiomic feature analysis. The proposed framework employs a hierarchical deep learning architecture designed to extract discriminative visual patterns associated with ulcerative regions from high-resolution gastric images. Convolutional neural networks are utilized to capture low-level spatial and textural cues, while deeper layers learn high-level semantic representations of mucosal abnormalities. In parallel, a comprehensive set of handcrafted radiomic features is extracted across multiple spatial scales to quantify subtle intensity variations, texture heterogeneity, shape descriptors, and morphological irregularities that are often overlooked by purely data-driven models. To effectively exploit the complementary nature of deep and radiomic representations, an attention-guided feature fusion mechanism is introduced, enabling adaptive weighting of multi-scale features based on their diagnostic relevance. Extensive experimental evaluations are conducted using a curated dataset of high-resolution gastric endoscopic images, incorporating rigorous preprocessing, data augmentation, and cross-validation strategies. The proposed hybrid framework demonstrates superior diagnostic performance compared to conventional deep learning and radiomics-only baselines, achieving notable improvements in accuracy, sensitivity, specificity, and area under the receiver operating characteristic curve. Ablation studies further confirm the critical contribution of multi-scale radiomic features and attention-based fusion to overall system robustness and generalization capability. Beyond performance gains, the framework emphasizes interpretability by enabling visualization of salient image regions and dominant radiomic signatures contributing to model decisions, thereby enhancing clinical trust and transparency. The results suggest that the proposed AI-driven diagnostic system has strong potential for real-world clinical deployment as a decision-support tool, facilitating early gastric ulcer detection, reducing diagnostic variability, and ultimately improving patient outcomes. This work highlights the value of hybrid intelligence that combines deep learning with domain-specific radiomic analysis for next-generation gastrointestinal disease diagnosis.