Email: submit@sarpublication.com (24x7 Online Support)
South Asian Research Journal of Engineering and Technology (SARJET)
Volume-8 | Issue-05
Original Research Article
Capsule-Based Brain Tumor Classification with Residual and Pyramid Spatial Features
Mustafa Challoob Albazoone*
Published : Nov. 20, 2026
DOI : https://doi.org/10.36346/sarjet.2026.v08i05.003
Abstract
Accurate classification of brain tumors from Magnetic Resonance Imaging (MRI) may help with the clinical evaluation process. However, many automated classification systems have demonstrated sensitivity to variation in tumor appearance, spatial relationships and the distribution of the training data. This study evaluates the classification performance of capsules in three different configurations: the baseline HitNet configuration, a Res-HitNet configuration with residual connections, and a PSP-HitNet configuration with hierarchical spatial analysis. All of these configurations use a Hit-or-Miss (HoM) capsule layer that maps each class-specific capsule vector to the extracted image features. The Res-HitNet configuration utilizes shortcut connections that allow for integration of features throughout the network. Processing stages, while PSP-HitNet incorporates multi-scale branches with different spatial filter sizes to represent patterns at more than one scale. The best-performing architecture is then evaluated in an uncertainty-guided deep active-learning setting, in which low-confidence unlabeled examples are selected for annotation and added incrementally to the labeled training set. The paper reports experiments on 3,264 MRI images carried out using a Kaggle dataset. The dataset characterization lists 926 glioma images, 937 meningioma images, 901 pituitary-tumor images, and 500 images without a tumor. Five-fold cross-validation is reported for the main model comparisons. The paper reports a best overall accuracy of 95.98% for the proposed approach and describes faster improvement from uncertainty-based sample selection than from random selection during early active-learning rounds. Additionally, this study demonstrated that the decision verification performance varied from approximately 89.3% to 93% when the capsule-based classification system underwent a longer period of reinforcement learning. The findings suggest that capsule-based classification, along with multi-scale spatial processing, has potential as an effective method for evaluating the performance of capsule-based systems.

About Us


South Asian Research Publication (SAR Publication) is a publisher for scientific online and print journals started with collaboration with other scientific organizations, institutions, academicians and researchers. SAR Publication is keen to make itself as a leading publisher for scientific and academic journals with quality peer review and rapid publication... Read More Here

Copyright © SAR Publication, All Rights Reserved

Developed by JM