Year: 2025
Type: article
Abstract: —The manual delineation of brain neoplasms using magnetic resonance imaging (MRI) is complex, labor-intensive, and time-consuming. Accurate segmentation of brain tumors is critical for neuro-oncology diagnosis, radiation planning, and evaluating treatment response. Traditional automated segmentation methods rely on handcrafted feature extraction pipelines, which often lack generalizability. Moreover, standard deep learning approaches, especially convolutional neural networks (CNNs), require large annotated datasets for supervised learning—data that are scarce and costly in clinical neuroimaging. To address these challenges, we propose a novel dual-pathway Group Convolutional Neural Network (Group-CNN) architecture tailored for brain tumor segmentation. This design simultaneously captures fine-grained local spatial features and globally relevant representations through multi-scale receptive fields. The model includes a bidirectional CNN similarity mechanism, reducing training instability and overfitting by optimizing shared parameters across symmetric network paths. Additionally, a cascaded architecture is integrated into a two-branch multicast CNN topology, where outputs from the base CNN serve as auxiliary priors and are aggregated through hierarchical fusion at the final layers. Validation on benchmark datasets BRATS2013 and BRATS2015 demonstrates that the proposed Group-CNN within the dual-pathway framework achieves superior segmentation accuracy and generalization, outperforming state-of-the-art baselines while maintaining computational efficiency. Keywords — Brain Tumor Segmentation, MRI, Deep Learning, Convolutional Neural Networks (CNNs),Dual-Pathway Architecture, Group-CNN, Multi-scale Features, Bidirectional Similarity Network, Cascaded Architecture, BRATS Dataset.
DOI: 10.55041/ijsrem47854
Author: Dr Arun Chakravarthy R
Language: English
Subfield: Neurology
Field: Neuroscience
Domain: Life Sciences
Open Access status: bronze
OpenAlex ID: W4410360903