ASA-CoroNet: Adaptive Self-attention Network for COVID-19 Automated Diagnosis using Chest X-ray Images

Fujun Wu, Jianjun Yuan, Yuxi Li, Jinyi Li, Ming Ye
Proceedings of the 1st Workshop on Healthcare AI and COVID-19, ICML 2022, PMLR 184:11-20, 2022.

Abstract

Computer-assisted imagery analysis based on chest X-ray images plays a crucial role in the clinical diagnosis and screening of COVID-19. However, the radiographic features of chest X-rays are highly complex and irregular in shape. Moreover, the size and location of the lesion regions vary greatly with infection stages and patients, thus dramatically increasing the difficulty of COVID-19 identification. A lightweight adaptive self-attention network is developed to address this problem, namely ASA-CoroNet. It firstly extracts underlying features using a depthwise separable convolution-based backbone, then further identifies lesion regions through an adaptive self-attentive module, and finally utilizes a homogeneous vector capsule layer to map the obtained features into capsule vectors to instantiate detection objects accurately. Extensive experimental results demonstrate that the proposed model outperforms the state-of-the-art methods and obtains competitive results on limited datasets. More importantly, the trainable params of the proposed model are reduced by 7x compared to the state-of-the-art capsule network. In addition, we also interpret the proposed model using different class activation techniques and confirm the validity of the three components through numerous ablation studies.

Cite this Paper

BibTeX
@InProceedings{pmlr-v184-wu22a, title = {ASA-CoroNet: Adaptive Self-attention Network for COVID-19 Automated Diagnosis using Chest X-ray Images}, author = {Wu, Fujun and Yuan, Jianjun and Li, Yuxi and Li, Jinyi and Ye, Ming}, booktitle = {Proceedings of the 1st Workshop on Healthcare AI and COVID-19, ICML 2022}, pages = {11--20}, year = {2022}, editor = {Xu, Peng and Zhu, Tingting and Zhu, Pengkai and Clifton, David A. and Belgrave, Danielle and Zhang, Yuanting}, volume = {184}, series = {Proceedings of Machine Learning Research}, month = {22 Jul}, publisher = {PMLR}, pdf = {https://proceedings.mlr.press/v184/wu22a/wu22a.pdf}, url = {https://proceedings.mlr.press/v184/wu22a.html}, abstract = {Computer-assisted imagery analysis based on chest X-ray images plays a crucial role in the clinical diagnosis and screening of COVID-19. However, the radiographic features of chest X-rays are highly complex and irregular in shape. Moreover, the size and location of the lesion regions vary greatly with infection stages and patients, thus dramatically increasing the difficulty of COVID-19 identification. A lightweight adaptive self-attention network is developed to address this problem, namely ASA-CoroNet. It firstly extracts underlying features using a depthwise separable convolution-based backbone, then further identifies lesion regions through an adaptive self-attentive module, and finally utilizes a homogeneous vector capsule layer to map the obtained features into capsule vectors to instantiate detection objects accurately. Extensive experimental results demonstrate that the proposed model outperforms the state-of-the-art methods and obtains competitive results on limited datasets. More importantly, the trainable params of the proposed model are reduced by 7x compared to the state-of-the-art capsule network. In addition, we also interpret the proposed model using different class activation techniques and confirm the validity of the three components through numerous ablation studies.} }
Endnote
%0 Conference Paper %T ASA-CoroNet: Adaptive Self-attention Network for COVID-19 Automated Diagnosis using Chest X-ray Images %A Fujun Wu %A Jianjun Yuan %A Yuxi Li %A Jinyi Li %A Ming Ye %B Proceedings of the 1st Workshop on Healthcare AI and COVID-19, ICML 2022 %C Proceedings of Machine Learning Research %D 2022 %E Peng Xu %E Tingting Zhu %E Pengkai Zhu %E David A. Clifton %E Danielle Belgrave %E Yuanting Zhang %F pmlr-v184-wu22a %I PMLR %P 11--20 %U https://proceedings.mlr.press/v184/wu22a.html %V 184 %X Computer-assisted imagery analysis based on chest X-ray images plays a crucial role in the clinical diagnosis and screening of COVID-19. However, the radiographic features of chest X-rays are highly complex and irregular in shape. Moreover, the size and location of the lesion regions vary greatly with infection stages and patients, thus dramatically increasing the difficulty of COVID-19 identification. A lightweight adaptive self-attention network is developed to address this problem, namely ASA-CoroNet. It firstly extracts underlying features using a depthwise separable convolution-based backbone, then further identifies lesion regions through an adaptive self-attentive module, and finally utilizes a homogeneous vector capsule layer to map the obtained features into capsule vectors to instantiate detection objects accurately. Extensive experimental results demonstrate that the proposed model outperforms the state-of-the-art methods and obtains competitive results on limited datasets. More importantly, the trainable params of the proposed model are reduced by 7x compared to the state-of-the-art capsule network. In addition, we also interpret the proposed model using different class activation techniques and confirm the validity of the three components through numerous ablation studies.
APA
Wu, F., Yuan, J., Li, Y., Li, J. & Ye, M.. (2022). ASA-CoroNet: Adaptive Self-attention Network for COVID-19 Automated Diagnosis using Chest X-ray Images. Proceedings of the 1st Workshop on Healthcare AI and COVID-19, ICML 2022, in Proceedings of Machine Learning Research 184:11-20 Available from https://proceedings.mlr.press/v184/wu22a.html.

Related Material