Back to all papers

Dynamic attention ordering in self-adaptive convolutional block attention module for chest X-ray image classification.

November 17, 2026pubmed logopapers

Authors

Shaari FN,Abdul Nasir AS,Wan Ahmed WA

Affiliations (2)

  • Faculty of Electrical Engineering & Technology, Universiti Malaysia Perlis, Pauh Putra Campus, Arau, 02600, Malaysia.
  • Universiti Sains Malaysia - Kampus Kesihatan, Kubang Kerian, Kubang Kerian, Kelantan, 16150, Malaysia.

Abstract

Convolutional Block Attention Module (CBAM) is widely used attention mechanism to boost the feature representation ability of the Convolutional Neural Network (CNN) by performing channel attention and spatial attention alternately. However, CBAM is hard to use for various distributions of the features, because of its fixed channel-to-spatial attention ordering. Therefore, this study proposes a Self-Adaptive Convolutional Block Attention Module (SA-CBAM) which introduces a lightweight auxiliary decision network for dynamically ordering the attention. The auxiliary network uses the global characteristics of the attention feature map to determine whether the spatial attention or the channel attention should be applied first, which is more flexible than the conventional CBAM for feature refinement. The effectiveness of the proposed module is assessed by systematically investigating six variants of the SA-CBAM where the standard, lightweight, early-stage, middle-stage and late-stage are investigated using a baseline CNN. The top-performing variant was then compared to representative attentional approaches and shown to be valid on several benchmark datasets that included chest X-ray (CXR) classification, remote sensing, natural image, microscopic image, and histopathological image classification tasks. Finally, the optimal SA-CBAM was combined with a hybrid CNN-LSTM and evaluated for sequential feature learning. Through experimental results, it was verified that the proposed SA-CBAM outperformed the conventional CBAM on all the test sets with similar complexity. The CNN+SA-CBAM model had accuracy of 90.88% and MCC of 86.52% on the primary CXR dataset, while the hybrid CNN+SA-CBAM-LSTM model further improved the accuracy to 99.88%. The results on different datasets show the effectiveness and computational efficiency of the proposed dynamic attention-ordering mechanism and make SA-CBAM robust and generalizable attention module for deep image classification.

Topics

Journal Article

Ready to Sharpen Your Edge?

Subscribe to join 11k+ peers who rely on RadAISlice. Get the essential weekly briefing that empowers you to navigate the future of radiology.

We respect your privacy. Unsubscribe at any time.