CycleGAN network for photoacoustic histology

An unsupervised deep learning algorithm based on cycle-consistent generative adversarial networks (CycleGAN ) converted the UV-PAM images into H&E-like pseudocolor histologic images, allowing the pathologists to readily identify the cancerous features following existing pattern-recognition parameters. Unlike supervised deep learning methods such as generational adversarial networks (GAN), the unsupervised deep learning method based on CycleGAN does not require coupled pairs of stained and unstained images. It avoids the need for well-aligned UV-PAM and H&E-stained images for neural network training, which can be challenging to acquire due to artifacts caused by sample preparation-induced morphology changes.

基于循环一致性生成对抗网络（CycleGAN）的无监督深度学习算法，可将紫外光声显微镜（UV-PAM）图像转换为类苏木精-伊红（H&E）伪彩色组织学图像，使病理学家能够依托现有模式识别参数轻松识别癌变特征。与生成对抗网络（GAN）等监督深度学习方法不同，基于CycleGAN的无监督深度学习算法无需配对的染色与未染色图像样本，同时规避了为神经网络训练准备精准对齐的UV-PAM与H&E染色图像的需求——由于样本制备引发的形态变化会产生伪影，这类图像的获取难度极大。