Аннотация:Histological images serve as crucial tools in the diagnosis and treatment of diverse afflictions. However, the acquisition of images exhibiting exceptional resolution whole slide images, WSIs, capturing intricate textures and vital nuances, can present a formidable challenge, primarily due to the requirement of expensive and intricate apparatus, proficient personnel, and considerable time commitments. To tackle this predicament, it is important that we conceive an effective and precise framework to increase the resolution of whole slide histological images. There are several algorithms used for super-resolution, including interpolation-based, deep learning based and bayes based algorithms. After scrutinizing and dissecting the available super-resolution models and algorithms, we arrived at the conclusion that the most suitable approach for histological WSIs would be to fine-tune the already trained Real-ESRGAN model to reconstruct histological images and apply it in a patch-based way. For histological WSIs, it is typical to have a lagre number of empty areas that do not contain tissue. To circumvent the impact of these empty areas on the model’s efficiency, we filter the dataset using Shannon’s information entropy. Furthermore, we have modified the structure of the loss function to optimize the reconstruction of low-level details of histological images. In this study, we fine-tune the pre-trained Real-ESRGAN model using the histological image dataset PATH-DT-MSU. It enabled us to outperform all preexisting models in terms of reconstructing low-level details in WSI histological images. Moreover, without retraining the model, we have tested it on additional histological image datasets, thereby proving its high generalization ability.
