| Summary: | Visual attribution (VA) methods play a crucial role in tuberculosis (TB) research by providing valuable
insights into disease patterns and aiding in diagnostic interpretation. The advent of generative adversarial
network (GAN)-based VA methods has gained significant attention from researchers due to their ability to
generate fine-grained feature maps that accurately reflect the location of lesions. These methods localize
lesions by converting chest X-ray (CXR) images containing lesions into normal CXR images and analyzing
the differences between the two. However, current methods only perform surface-level transformations,
neglecting the vital information of whether lesions are present. Moreover, the transformation process assigns
equal weights to the entire image, without specifically prioritizing the regions with a higher probability of
lesions occurrence. In this study, a novel framework is proposed, namely the class activation mapping-guided
tuberculosis visual attribution generative adversarial network (TBVA-GAN). This innovative model
leverages the informative regions derived from class activation mapping to effectively guide the GAN in
prioritizing the transformation of these crucial areas. Moreover, to guarantee the precision of TB localization,
an auxiliary TB detection model is incorporated, ensuring that the converted CXR images are devoid of TB
pathology. By employing this additional verification step, the accuracy of TB localization is significantly
enhanced. The proposed TBVA-GAN in this study achieves promising VA results on the TBX11K dataset,
surpassing existing GAN-based TB VA models.
|
|---|