A Resource-Efficient Deep Learning Framework for Gastrointestinal Image Classification

Abstract

In this paper, we propose a cost-effective deep learning solution for diagnosing GI diseases, focusing on GERD and colorectal polyps, through endoscopic images. We explore the problems of executing high-performance AI models in resource-restricted clinical settings through incorporating transfer learning, ensemble learning, and a new lightweight architecture named MiniMedNet (~32.5k parameters). The dataset is a global dataset and includes four classes (GERD, GERD Normal, Polyp, Polyp Normal) which are collected from one publicly available Mendeley repository and heavily pre-processed (augmentation, resizing, normalization). Several off-the shelf models (EfficientNetB3, ResNet50, DenseNet121, and MobileNetV2) were tested individually and in ensemble for the purpose of setting the baselines. We show that MiniMedNet, a network model designed from the ground up to use the fewest number of parameters possible, attains a test accuracy of 78\% on average, which is comparable to the performance of other more compute-laden models. Exhaustive analysis with metrics accuracy, precision, recall, F1-score and Grad-CAM visualizations reiterate the ability of the model to retain interpretatibility and reliability. Our approach presents a potential route for deploying AI-supported diagnosis in resource- limited healthcare, by trading-off between diagnostic performance and computational cost, and by enabling scalable, clinically practical AI deployment.