Modified CBAM: sub-block pooling for improved channel and spatial attention.

Hussaini, Hamza; Bano, Shahana; Elyan, Eyad; Moreno-Garcia, Carlos Francisco

Abstract

The Convolutional Block Attention Module (CBAM) has emerged as a widely adopted attention mechanism, as it seamlessly integrates into the Convolutional Neural Network (CNN) architecture with minimal computational overhead. However, its reliance on global average and maximum pooling in the channel and spatial attention modules leads to information loss, particularly in scenarios demanding fine-grained feature analysis, such as medical imaging. In this paper, we propose the Modified CBAM (MCBAM) to address this critical limitation. This novel framework eliminates the dependence on global pooling by introducing a sub-block pooling strategy that captures nuanced feature relationships, preserving critical spatial and channel-wise information. MCBAM iteratively computes attention maps along channel and spatial dimensions, adaptively refining features for superior representational power. Comprehensive evaluations on diverse datasets, including C-NMC (acute lymphoblastic leukemia), PCB (peripheral blood cells), and COVID-19 (Chest X-ray), demonstrate the efficacy of MCBAM. Additionally, we evaluate MCBAM against similar alternatives, such as the Bottleneck Attention Module (BAM), Normalisation-Based Attention Module (NAM), and Triplet Attention Module (TAM), demonstrating that MCBAM consistently outperforms these advanced attention mechanisms across all datasets and metrics. Furthermore, results reveal that MCBAM surpasses the standard CBAM and establishes itself as a robust and effective enhancement for attention mechanisms, with notable improvements in medical imaging tasks, offering critical advantages in complex scenarios.

Citation

HUSSAINI, H., BANO, S., ELYAN, E. and MORENO-GARCIA, C.F. 2026. Modified CBAM: sub-block pooling for improved channel and spatial attention. In Ali, S., Hogg, D.C. and Peckham, M. (eds.) Proceedings of the 29th Annual conference of the Medical image understanding and analysis 2025 (MIUA 2025), 15-17 July 2025, Leeds, UK. Lecture notes in computer science, 15917. Cham: Springer [online], part II, pages 116-130. Available from: https://doi.org/10.1007/978-3-031-98691-8_9

• Presentation Conference Type: Conference Paper (published)
• Conference Name: 29th Annual conference of the Medical image understanding and analysis 2025 (MIUA 2025)
• Start Date: Jul 15, 2025
• End Date: Jul 17, 2025
• Online Publication Date: Jul 15, 2025
• Publication Date: Jan 1, 2026
• Deposit Date: Jul 18, 2025
• Publicly Available Date: Jul 16, 2026
• Publisher: Springer
• Peer Reviewed: Peer Reviewed
• Pages: 116-130
• Series Title: Lecture Notes in Computer Science
• Series Number: 15917
• ISBN: 9783031986901
• DOI: https://doi.org/10.1007/978-3-031-98691-8_9
• Keywords: Modified convolutional block attention module (MCBAM); Convolutional block attention module (CBAM); Convolutional neural network (CNN); Attention mechanism; Medical imaging; Global pooling; Sub-block pooling; Channel attention; Spatial attention
• Public URL: https://rgu-repository.worktribe.com/output/2929216
• Additional Information: The code for the CNN-MCBAM can be accessed via: https://github.com/fatimza-2013/brainiac.