Faculty Publications
Permanent URI for this communityhttps://idr.nitk.ac.in/handle/123456789/18736
Publications by NITK Faculty
Browse
Search Results
Item Multi-class Classification of Wireless Capsule Endoscopy Images By Using the Fusion of Pre-trained Networks and Fusion Residual Block(Springer Science and Business Media Deutschland GmbH, 2025) Lavanya, K.; Aparna., P.In order to identify diverse endoscopic images of gastrointestinal (GI) problems, this study presented a multi-fused residual convolutional neural network (MFuRe-CNN) with auxiliary fusion layers (AuxFL) and a fusion residual block (FuRB) with alpha dropouts (ADO). The implemented MFuRe-CNN dealt with five instances sourced from reputed databases such as the KVASIR database, the ETIS-Larib Polyp Database, and the Red Lesion Endoscopic Database, which included Esophagitis, Ulcerative colitis, Polyps, Healthy colon and Bleeding images. The proposed model was created by fusing three cutting-edge models fused into a single-feature extraction pipeline with its layers that are truncated and partially frozen. This model which despite using a small portion of the computing power of most current cutting-edge models, helped spread robust features and enhanced diagnostic performance. In addition, compared to those without the aforementioned components, the MFuRe-CNN in conjunction with AuxFLs, DOs and FuRB has significantly reduced overfitting and performance saturation. With merely 4.8 million parameters, this model achieved test accuracy of 93.79 percent, outperforming the majority of DCNNs that are trained in the traditional manner. Thus, the suggested MFuRe-CNN could improve GI tract diagnosis more cheaply than ensembles and outperform other traditional pre-trained as well as fine-tuned Deep convolutional neural networks. © The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2025.