Super-resolution reconstruction of polarized cancerous tissue images using a multi-attention residual generative adversarial network

To accurately distinguish cancerous from non-cancerous regions in polarized images of cancerous tissues, we propose a deep learning–based super-resolution (SR) reconstruction method. Our approach utilizes a multi-attention residual generative adversarial network that incorporates channel, spatial, and multi-scale attention mechanisms within a residual network framework. This network design allows for the extraction of features from multiple scales, enriching the details in the reconstructed images. By comparing our method with traditional interpolation methods and other deep learning–based SR techniques, we demonstrate significant improvements in both objective metrics (peak signal-to-noise ratio increased by up to 4.51 dB and structural similarity index measure increased by up to 0.06) and visual quality. Our approach produces images with clearer texture details and better human visual effects, providing valuable data support for the study of polarized images of cancerous tissues.