Bidirectional Motion-Temporal Deep Learning for Explainable Multi-Class Classification of Gastrointestinal Lesions in Wireless Capsule Endoscopy

Sarfaraz Natha^1,*, Mohammad Siraj^2,*, Mohammed Muflih Alamer³, Aaqid Syed⁴, Ayesha Shafique⁵, Kashan Memon⁶
1 Department of Software Engineering, Sir Syed University of Engineering & Technology, Karachi, Pakistan
2 Department of Electrical Engineering, College of Engineering, King Saud University, Riyadh, Saudi Arabia
3 Department of Curriculum and Instruction, College of Education, King Saud University, Riyadh, Saudi Arabia
4 Resident, Internal Medicine, Mobile Infirmary Medical Center, Mobile, AL, USA
5 School of IoT Engineering, Wuxi Taihu University, Jiangsu Key (Construction) Laboratory of Intelligent IoT Technology and Applications in Universities, Wuxi, China
6 Department of Electronic Engineering and Information Sciences, University of Science and Technology of China, Hefei, China
* Corresponding Author: Sarfaraz Natha. Email: email ; Mohammad Siraj. Email: email
(This article belongs to the Special Issue: Artificial Intelligence in Healthcare: Current Challenges, Emerging Trends, and Future Directions)

Computer Modeling in Engineering & Sciences https://doi.org/10.32604/cmes.2026.087211

Received 12 June 2026; Accepted 27 August 2026; Published online 11 September 2026

Download PDF

Abstract

Gastrointestinal (GI) tract cancers are a serious health concern worldwide due to their high mortality rates. Wireless Capsule Endoscopy (WCE) provides a valuable non-invasive approach for detecting gastrointestinal abnormalities that may be associated with cancer. Despite WCE examinations generating many images, manual assessment is time-consuming. Therefore, automated methods capable of accurate and efficient lesion classification are highly desirable. Deep Learning (DL) techniques have demonstrated considerable potential for medical image analysis. However, many existing deep learning methods struggle to capture both broader contextual relationships and suitable patterns at the same time. While many are limited to binary classification. To address this limitation, this study proposed Bidirectional Motion Temporal (BiMT) for multi-class classification of gastrointestinal images by exploiting spatial and temporal information from image sequences. BiMT architecture effectively integrates spatial and temporal information to reduce the limitation of existing approaches in medical image analysis. In the first stage, pretrained CNN models, including InceptionNetV3 and DenseNet201, are employed to extract discriminative feature representations from the WCE dataset. This feature extraction process captures rich spatial characteristics from WCE images, providing a robust foundation for subsequent analysis. In the second stage, a BiLSTM model is used to capture the temporal dependencies between consecutive frames and is combined with Multi-Head Self-Attention (MHSA) to effectively model short-term contextual relationships. Furthermore, a custom Transformer encoder incorporating relative positional embeddings to model sequential information and increase gastrointestinal disease detection. During training, Categorical Focal Loss (CFL) is employed to emphasize difficult-to-classify samples and clinically relevant features. We combine the publicly available Kvasir-Capsule-v1 and Kvasir-v2 datasets to construct a unified dataset comprising four clinically relevant classes: ulcerative colitis, polyps, dyed-lifted polyps, and normal. The more sophisticated ConvNeXt-BiMT model achieved an average accuracy of 98.41%.