Cross prior Bayesian attention with correlated inception and residual learning for brain tumor classification using...

Brain tumor classification from magnetic resonance (MR) images is crucial for early diagnosis and effective treatment planning. However, the homogeneity of tumors across different categories poses a challenge. Although, attention-based convolutional neural networks (CNNs) approaches have shown promising results in brain tumor classification, simultaneous consideration of both spatial and channel-specific features remains limited.

This study proposes a novel model that integrates Bi-FocusNet with correlated learning and CB-Attention. Bi-FocusNet is designed to concentrate on both spatial and channel-wise tumor features by using two complementary learning methodologies: correlated spatial inception learning and correlated channel residual learning. These learnings extract richer and more diverse feature representations from tumor lesions of varied sizes, significantly enhancing the model's learning capacity. The CB-Attention mechanism works as a cross-learning module, facilitating interaction between the two learning methods to capture the missing information across spatial and channel-wise features.

Ablation studies and experiments were conducted using the BT-large-2c, Figshare, and Kaggle datasets. The proposed model outperformed existing classification methods in accuracy and other metrics, demonstrating enhanced performance on all three datasets with accuracies of 99.02 %, 97.06 %, and 96.44 %, respectively. Additionally, the BT-Merged-4c dataset was used to assess the ability to handle class variation, and 96.28 % accuracy was achieved.

The CB-CIRL Net improves the extraction of spatial and channel-wise features through the utilization of Bi-FocusNet with correlated learning and CB-Attention, resulting in enhanced classification accuracy across various datasets. The model's outstanding performance demonstrates its capacity to improve brain tumor diagnosis and clinical application.