A semantically robust adversarial training domain adaptation method for tuberculosis diagnosis

Domain adaptation is crucial for tuberculosis diagnosis from X-ray images, as it mitigates variations arising from different imaging conditions across hospitals, thereby enhancing the generalization capability and reliability of tuberculosis diagnostic models. Cycle-consistent adversarial training methods have shown promising results in medical image domain adaptation by learning domain-invariant feature representations. However, these approaches often fail to preserve critical diagnostic information, such as the presence of lesions, during the image translation process, leading to suboptimal diagnostic performance. To address this limitation, a semantically robust adversarial training domain adaptation (SRAT-DA) method is proposed that maintains semantic consistency while translating image styles across domains. Cycle-consistent adversarial training is extended by incorporating semantic constraints to retain clinically relevant information during image transformation. The approach is evaluated on domain adaptation tasks between the Shenzhen and TBX11K TB X-ray datasets. The promising results, achieving an average accuracy of 73.7%, sensitivity of 73.0%, and specificity of 74.1%, demonstrate the effectiveness in improving TB diagnosis under domain shift scenarios.

Text 121499.pdf - Published Version Restricted to Repository staff only Download (557kB) Official URL or Download Paper: https://ieeexplore.ieee.org/document/10788597/

Actions (login required)

View Item