Evaluación de Redes Adversariales Antagónicas (GANs) para eliminar el ruido de imágenes ultrasónicas de mama

Abstract

Abstract: Ultrasound plays an important role in the early detection and diagnosis of breast cancer. However, speckle noise affects this type of medical imaging, degrading its visual radiological evaluation and causing a series of difficulties when to identify malignant and benign regions. Noise removal is an important step in the pre-processing of medical images, since it restores the maximum details and the image information, achieving successful precision in segmentation and anomaly classification. To reduce speckle noise and preserve features of the ultrasound image in the breast, we proposed two GAN models for the elimination of the noise: (i) conditional GAN ​​(CGAN) and (ii) Wassertein GAN (WGAN). The relationship metrics peak signal-to-noise ratio (PSNR) and the structural similarity index (SSIM) allowed a Statistical evaluation of image quality after denoising. The analysis Experimental study shows that the CGAN method achieves a higher image quality in terms of PSNR of 38.18 dB and SSIM of 0.96, higher than WGAN, PSNR of 33.0068 dB and SSIM of 0.919955. As a conclusion, the CGAN network shows higher quality values images in the BUSI training data sets, these results are by the robustness of the model itself due to the use of the U-Net architecture in the network generator and PatchGAN for the discriminator. Therefore, we confirm that GAN networks can assist in the elimination of noise from medical ultrasound images, which in the future would serve as input to improve the accuracy of the segmentation and classification of the breast imaging in a computer assisted system to help radiologists improve breast cancer detection.