Multi-scale self-attention generative adversarial network for pathology image restoration

High-quality histopathology images are significant for accurate diagnosis and symptomatic treatment. However, local cross-contamination or missing data are common phenomena due to many factors, such as the superposition of foreign bodies and improper operations in obtaining and processing pathological digital images. The interpretation of such images is time-consuming, laborious, and inaccurate. Thus, it is necessary to improve diagnosis accuracy by reconstructing pathological images. However, corrupted image restoration is a challenging task, especially for pathological images. Therefore, we propose a multi-scale self-attention generative adversarial network (MSSA GAN) to restore colon tissue pathological images. The MSSA GAN uses a self-attention mechanism in the generator to efficiently learn the correlations between the corrupted and uncorrupted areas at multiple scales. After jointly optimizing the loss function and understanding the semantic features of pathology images, the network guides the generator in these scales to generate restored pathological images with precise details. The results demonstrated that the proposed method could obtain pixel-level photorealism for histopathology images. Parameters such as RMSE, PSNR, and SSIM of the restored image reached 2.094, 41.96 dB, and 0.9979, respectively. Qualitative and quantitative comparisons with other restoration approaches illustrate the superior performance of the improved algorithm for pathological image restoration.