Published Online:October 2025
Product Name:The IUP Journal of Computer Sciences
Product Type:Article
Product Code:IJCS021025
DOI:10.71329/IUPJCS/2025.19.4.17-42
Author Name:Earnest Ebenezer, Ashli Paul, Chaitanya V Mahamuni, Christopher Aseer J Albert, Sophia John Chavakula and Ann Rachel Koshy
Availability:YES
Subject/Domain:Engineering
Download Format:PDF
Pages:17-42
The paper shows how capsule networks (CapsNets) alleviate the innate shortcomings of convolutional neural networks (CNNs) by preserving the part-whole relationship that results in improvement of the effectiveness of segmentation. In particular, it tests CapsNets against several benchmark datasets—brain tumors (BraTS), skin lesions (ISIC) and lung nodules (LIDC-IDRI)—demonstrating its generalizability to MRI, dermoscopic and X-ray images. Besides the empirical analysis, the paper also provides a personalized lightweight variation of the CapsNet model, which is specifically aimed at resource-constrained settings. The paper focuses on routing mechanism comparisons (i.e., dynamic routing vs. EM routing), and also on integration with transformer-based architecture and self-supervised learning to perform better feature extraction in limited labeled data environment. It also offers future directions, which include development of low-latency routing algorithms and hybrid capsule-transformer models and additions in AI interpretability frameworks to ensure clinical use. The findings will guide clinicians and AI practitioners in implementing CapsNets in reallife clinical practice at high precision and interpretability levels in medical image segmentation.
With artificial intelligence (AI), the interpretation of medical images can now actually be viewed automatically and accurately without the assistance of a human being.