Endoscopic Ultrasound Image Classification Using a Mobilenet-Resnet Distillation Model

Abstract

Accurate detection of gastrointestinal (GI) diseases such as adenocarcinomas through Endoscopic Ultrasound (EUS) imaging is vital for early diagnosis and improved clinical outcomes. However, the inherent noise, low contrast, and complex textures in EUS images pose major challenges for reliable automated analysis. This study presents AUTOEUS, an enhanced lightweight deep learning framework capable of both anatomical region classification and adenocarcinoma detection from EUS images. The proposed pipeline incorporates a two-stage preprocessing process using median filtering for noise suppression and Y-channel histogram equalization for contrast enhancement, significantly improving image clarity. A teacher–student knowledge distillation architecture is implemented, where a ResNet-50–based teacher network guides a compact convolutional student model, ensuring high accuracy with reduced computational cost. Experimental evaluation performed in MATLAB using augmented image datastores and five-fold binary classification metrics demonstrates strong diagnostic performance, achieving accuracies of 90.70% (cecum), 95.81% (ileum), 80.00% (pylorus), 90.23% (rectum), and 90.23% (stomach). Corresponding F1-scores reached up to 97.06%, validating the model’s strong precision–recall balance across multiple adenocarcinoma types. The visualization results confirm the framework’s ability to correctly classify true positive disease cases with high confidence. Owing to its lightweight architecture and robustness, AUTOEUS demonstrates high potential for real-time EUS-based disease detection and clinical decision support, offering a scalable, low-cost, and reliable tool for gastrointestinal cancer diagnosis.