Ultrasound Breast images denoising using Generative Adversarial Networks (GANs)

Ultrasound imaging plays an important role in screening early detection and breast cancer diagnosis. However, speckle noise affects medical ultrasound images and degrades the visual radiologic evaluation, which generally causes several difficulties in identifying the malignant and benignant regions. Denoising is an important step in preprocessing medical images, because it restores the maximum details preserving edges and all information of the images, achieving successful accuracy in anomalies classification. To reduce the speckle noise and retain image features well, we proposed two GANs models to breast ultrasound speckle denoising as preprocessing image, (i) Conditional GAN and (ii) WGAN. The better denoising image quality was measured by peak signal to noise ratio (PSNR) and structural similarity index (SSIM). The experimental analysis clearly shows that the CGAN method achieves better visual image quality in terms of PSNR=38.18 dB and SSIM= 0.96 with respect to WGAN model (PSNR=33.0068 dB and SSIM=0.91) on the small Ultrasound training datasets. Thus, we conclude that GANs can help in denoising ultrasound medical imaging, and as a future work these data can be used as computer system input for image segmentation and classification, reducing the hand-dependence and helping radiologists to improve breast cancer detection.超声成像在乳腺癌早期筛查与诊断中具有重要应用价值。然而，斑点噪声会对医学超声图像产生干扰，削弱其放射学视觉评估效果，进而为恶性与良性病灶区域的识别带来诸多挑战。去噪是医学图像预处理的关键步骤，其核心在于在保留图像边缘与全部细节信息的前提下还原图像，从而实现异常分类任务的高精度表现。为有效抑制斑点噪声并完整保留图像特征，本文提出两类用于乳腺超声斑点去噪的生成对抗网络（Generative Adversarial Networks，GANs）模型作为图像预处理方案，分别为条件生成对抗网络（Conditional GAN，CGAN）与瓦瑟斯坦生成对抗网络（Wasserstein GAN，WGAN）。本研究采用峰值信噪比（Peak Signal to Noise Ratio，PSNR）与结构相似性指数（Structural Similarity Index，SSIM）作为图像去噪质量的量化评估指标。基于小型乳腺超声训练数据集的实验分析表明，相较于WGAN模型（PSNR=33.0068 dB，SSIM=0.91），CGAN方法可获得更优的视觉图像质量，其PSNR可达38.18 dB，SSIM可达0.96。综上，本研究证实生成对抗网络可有效应用于医学超声图像去噪；未来可将此类去噪后图像作为计算机系统的输入，用于图像分割与分类任务，从而减少人工依赖，辅助放射科医师提升乳腺癌检出效能。