Diagnosis an‎d classification of gastrointestinal diseases based on endoscopic images using deep learning methods

3/21/2026 12:00:00 AM

In recent years, the application of deep learning techniques to the analysis of endoscopic images has played a significant role in the automated diagnosis of gastrointestinal diseases. Given the importance of early detection of gastrointestinal disorders, this study proposes an intelligent system based on deep neural networks for the identification of five common gastrointestinal diseases, including Barrett’s esophagus, upper esophagitis, lower polyps, lower ulcerative colitis, an‎d lower hemorrhoids. In this research, both labeled an‎d unlabeled images from the HyperKvasir dataset were used as the primary data source. The HyperKvasir dataset comprises a wide range of normal an‎d pathological regions of the upper an‎d lower gastrointestinal tract an‎d is considered one of the most comprehensive publicly available databases in this field. During the processing stage, labeled an‎d unlabeled data were simultaneously fed into two pre-trained networks. For labeled data, the Xception network was employed, while for unlabeled data, an autoencoder architecture consisting of an encoder an‎d a decoder was utilized. At this stage, both models extracted features in parallel, generating distinct representations of the textural an‎d structural characteristics of the images according to their respective architectures. In the proposed framework, all extracted features from both networks were fused an‎d provided as input to the final classification stage. Subsequently, after flattening the final convolutional layer, a fully connected layer with 64 neurons was added. These 64 neurons were then mapped to five output neurons, each corresponding to one anatomical location associated with the target diseases. A Softmax activation function was applied to compute the probability of each disease class. After training an‎d validation on the HyperKvasir dataset, the proposed model achieved high diagnostic performance in identifying the five target diseases. The final accuracies for Barrett’s esophagus, upper esophagitis, lower polyps, lower ulcerative colitis, an‎d lower hemorrhoids were 97%, 94%, 98%, 100%, an‎d 97%, respectively. These results demonstrate that the simultaneous integration of labeled an‎d unlabeled data, combined with deep architectures such as Xception an‎d autoencoders, significantly enhances the performance of automated gastrointestinal disease diagnosis. Ultimately, the proposed network accurately analyzes the structural an‎d textural features of endoscopic images, estimates the probability of each of the five diseases, an‎d can serve as an intelligent decision-support tool in clinical practice.

آندوسكوپي , HyperKvasir , يادگيري عميق , بيماري دستگاه گوارش

Deep Learning , Gastrointestinal Disease Diagnosis , Endoscopic Image Analysis , Semi-Supervised Learning , Feature Fusion , HyperKvasir Dataset

Narges Saber

Ahmad Ayatollahi