Artificial intelligence-powered tuberculosis detection with complementary domain attention model
Artificial intelligence-based X-ray image detection can significantly aid early tuberculosis (TB) detection. However, the varying distribution of X-ray image data across different hospitals has resulted in a decline in the model's performance when transitioning to a new dataset. Domain adaptati...
| Main Authors: | , , , , , |
|---|---|
| Format: | Article |
| Language: | English |
| Published: |
Elsevier B.V.
2025
|
| Online Access: | http://psasir.upm.edu.my/id/eprint/120525/ http://psasir.upm.edu.my/id/eprint/120525/1/120525.pdf |
| _version_ | 1848868196945231872 |
|---|---|
| author | Ding, Zeyu Yaakob, Razali Azman, Azreen Mohd Rum, Siti Nurulain Zakaria, Norfadhlina Ahmad Nazri, Azree Shahril |
| author_facet | Ding, Zeyu Yaakob, Razali Azman, Azreen Mohd Rum, Siti Nurulain Zakaria, Norfadhlina Ahmad Nazri, Azree Shahril |
| author_sort | Ding, Zeyu |
| building | UPM Institutional Repository |
| collection | Online Access |
| description | Artificial intelligence-based X-ray image detection can significantly aid early tuberculosis (TB) detection. However, the varying distribution of X-ray image data across different hospitals has resulted in a decline in the model's performance when transitioning to a new dataset. Domain adaptation techniques can effectively mitigate the impact of this issue. However, current domain adaptation methods align the entire image features between the source and target domains without explicitly focusing on regions containing transferable classification information across domains. Forced alignment of features across the entire image may lead to negative transfer. This paper proposes a complementary domain attention model (CDAM) for TB detection where the feature map is partitioned into domain-shared (DSH) and domain-specific (DSP) features. DSP features are complementary to DSH features. DSH features are responsible for mitigating the impact of domain gaps on classification. Consequently, they focus on areas containing classification information that can be transferred across domains. In contrast, the role of the DSP feature is to maximize the domain gap, concentrating its attention on areas rich in domain information. Given that the DSH and DSP features are complementary, when the DSP feature occupies domain-informative areas, it simultaneously encourages the DSH feature to focus more accurately on areas containing transferable classification information across domains, thereby enhancing classification performance. The objective of CDAM is to fully consider the importance of different regions within the feature map and mitigate negative transfer. The proposed method underwent domain adaptation experiments on the Shenzhen, Montgomery, and TBX11K datasets, achieving average accuracy, sensitivity, and specificity scores of 73.3%, 74.0%, and 72.7%, respectively. This result surpasses existing domain adaptation methods for TB data, providing evidence for the effectiveness and robustness of the proposed approach. |
| first_indexed | 2025-11-15T14:48:33Z |
| format | Article |
| id | upm-120525 |
| institution | Universiti Putra Malaysia |
| institution_category | Local University |
| language | English |
| last_indexed | 2025-11-15T14:48:33Z |
| publishDate | 2025 |
| publisher | Elsevier B.V. |
| recordtype | eprints |
| repository_type | Digital Repository |
| spelling | upm-1205252025-10-06T01:01:42Z http://psasir.upm.edu.my/id/eprint/120525/ Artificial intelligence-powered tuberculosis detection with complementary domain attention model Ding, Zeyu Yaakob, Razali Azman, Azreen Mohd Rum, Siti Nurulain Zakaria, Norfadhlina Ahmad Nazri, Azree Shahril Artificial intelligence-based X-ray image detection can significantly aid early tuberculosis (TB) detection. However, the varying distribution of X-ray image data across different hospitals has resulted in a decline in the model's performance when transitioning to a new dataset. Domain adaptation techniques can effectively mitigate the impact of this issue. However, current domain adaptation methods align the entire image features between the source and target domains without explicitly focusing on regions containing transferable classification information across domains. Forced alignment of features across the entire image may lead to negative transfer. This paper proposes a complementary domain attention model (CDAM) for TB detection where the feature map is partitioned into domain-shared (DSH) and domain-specific (DSP) features. DSP features are complementary to DSH features. DSH features are responsible for mitigating the impact of domain gaps on classification. Consequently, they focus on areas containing classification information that can be transferred across domains. In contrast, the role of the DSP feature is to maximize the domain gap, concentrating its attention on areas rich in domain information. Given that the DSH and DSP features are complementary, when the DSP feature occupies domain-informative areas, it simultaneously encourages the DSH feature to focus more accurately on areas containing transferable classification information across domains, thereby enhancing classification performance. The objective of CDAM is to fully consider the importance of different regions within the feature map and mitigate negative transfer. The proposed method underwent domain adaptation experiments on the Shenzhen, Montgomery, and TBX11K datasets, achieving average accuracy, sensitivity, and specificity scores of 73.3%, 74.0%, and 72.7%, respectively. This result surpasses existing domain adaptation methods for TB data, providing evidence for the effectiveness and robustness of the proposed approach. Elsevier B.V. 2025 Article PeerReviewed text en http://psasir.upm.edu.my/id/eprint/120525/1/120525.pdf Ding, Zeyu and Yaakob, Razali and Azman, Azreen and Mohd Rum, Siti Nurulain and Zakaria, Norfadhlina and Ahmad Nazri, Azree Shahril (2025) Artificial intelligence-powered tuberculosis detection with complementary domain attention model. Neurocomputing, 637. art. no. 130089. pp. 1-11. ISSN 0925-2312; eISSN: 1872-8286 https://www.sciencedirect.com/science/article/pii/S0925231225007611?via%3Dihub 10.1016/j.neucom.2025.130089 |
| spellingShingle | Ding, Zeyu Yaakob, Razali Azman, Azreen Mohd Rum, Siti Nurulain Zakaria, Norfadhlina Ahmad Nazri, Azree Shahril Artificial intelligence-powered tuberculosis detection with complementary domain attention model |
| title | Artificial intelligence-powered tuberculosis detection with complementary domain attention model |
| title_full | Artificial intelligence-powered tuberculosis detection with complementary domain attention model |
| title_fullStr | Artificial intelligence-powered tuberculosis detection with complementary domain attention model |
| title_full_unstemmed | Artificial intelligence-powered tuberculosis detection with complementary domain attention model |
| title_short | Artificial intelligence-powered tuberculosis detection with complementary domain attention model |
| title_sort | artificial intelligence-powered tuberculosis detection with complementary domain attention model |
| url | http://psasir.upm.edu.my/id/eprint/120525/ http://psasir.upm.edu.my/id/eprint/120525/ http://psasir.upm.edu.my/id/eprint/120525/ http://psasir.upm.edu.my/id/eprint/120525/1/120525.pdf |