Histopathological Image Super-resolution Using Attention Mechanisms

Histopathological images,which contain a large number of unevenly distributed and morphologically diverse cells,are currently the most important basis for cancer diagnosis and serve as a bridge between histopathology and computer image processing technology.Compared with the traditional method of examining a glass slider under the microscope,pathologists can observe histopathological images through digital terminals with higher clinical significance,which can help save time and improve diagnostic efficiency while reducing the subjective influence of histopathology diagnosis.With the support of artificial intelligence algorithms,computational histopathology can automate analysis of images for the purpose of aiding diagnosis and exploring deeper histopathological structures.High-quality histopathology images are the basis of relevant image processing techniques and need to be acquired with the help of advanced slide scanners.Limited by hardware equipment and microscopic imaging conditions,the acquisition of high-quality histopathological images is expensive and difficult to universalize.Super-resolution of histopathological images is rarely researched.The texture structure of histopathological images is more complex than natural images,which makes making super-resolution more difficult.In summary,there is an urgent need for research on super-resolution techniques for histopathological images.According to the characteristics of histopathological images,this paper focuses on the super-resolution techniques,and the research results achieved are as follows:Firstly,we propose a super-resolution method based on dynamic channel attention mechanism to improve the accuracy of super-resolution of histopathological images.The existing histopathology image super-resolution methods use all early methods of natural image super-resolution without introducing the latest image super-resolution mechanisms and without using real datasets.Attentional mechanism can focus on learning complex features such as texture details of histopathological images.The model proposed in this paper uses a dynamic channel attention mechanism to effectively extract high-frequency features of histopathological images,which can enhance the feature representation capability of the network.The skip-connection can combine the existing low-level information in the low-resolution space with the high-level information,which enables the network to focus on the reconstruction of high-frequency features.Experiments show that the method based on the dynamic channel attention mechanism outperforms other mainstream super-resolution methods overall on real histopathological image datasets.Secondly,we propose a super-resolution generative adversarial network based on a hybrid domain attention mechanism.Histopathological images,as the basis for doctors’ diagnosis,need to ensure the realistic degree of image details and visual effects after super-resolution reconstruction.The dynamic attention mechanism proposed above only considers the feature information between different channels and does not consider the importance of spatial location information.In this paper,we use the generative adversarial network,and the generator network learns the deep implicit features through a multiattention mechanism to make the generated image details more reliable.The training process makes the super-resolution image infinitely close to the real image by perceptual loss and adversarial loss.During the testing phase,only the generator network is retained for super-resolution reconstruction.The experimental results show that the method proposed in this paper not only further improves the objective metrics but also is closer to the natural high-resolution image in terms of visual effect.