Abstract
Medical hyperspectral imaging (MHSI) presents opportunities for computational pathology and precision medicine. However, the availability of medical hyperspectral image datasets for training is limited, and the existing enhancement methods generate low quality images, which limits the performance of early pathology recognition. Therefore, in this paper, we propose a high-quality medical hyperspectral image generation algorithm based on an improved recurrent generative adversarial network (Med-CycleGAN) to improve the performance of early medical pathology recognition. First, a module based on the encoder-decoder structure is introduced into the cyclic generative adversarial network (Cycle-GAN). This module combines self-attention layers. The self-attention layer effectively captures long-range dependencies in medical hyperspectral data, which helps to extract and retain important detail information in the image, thus enhancing the model's ability to accurately learn and generate medical features. Second, based on the loss correction principle, a new loss function, gradient penalized (WGAN-GP) loss, is embedded in the network to facilitate the balance between the generator and discriminator. Experimental results of our model on hyperspectral Choledoch Dataset confirm the effectiveness of the Med-CycleGAN approach, demonstrating its superiority and potential for medical hyperspectral image data enhancement.