CAM-Guided Spatially Denormalization for Lesion Weakly-Supervised Localization

Automatic localization of lesion regions is helpful to assist doctors in the diagnosis and treatment of diseases. In this paper, we propose a weakly-supervised 2D lung lesion localization method. Our method only requires image-level labels to perform cross-domain translation between normal and patient images through image style transfer. Finally realizes the conversion of the patient's image into the corresponding healthy person's image. The final lesion area is gained by comparing the difference between the patient image and the corresponding healthy version. On this basis, Class Activation Map (CAM) is used as additional semantic supervisory information incorporated into the generator through the SPADE layer to address the drawback of the current methods with blurred localization boundaries and noise in non-relevant areas. Experimental results on a real-world chest CT dataset demon-strate that the proposed method yields better performance than comparable methods.