作者:王紫旋，谢怡宁，何勇军
Authors:WANG Zi-xuan,XIE Yi-ning,HE Yong-jun摘要:摘要:病理辅助诊断系统采用在切片上放置滤光片的方法实现光源过滤，得到单色平行光，用于计算图像的光密度参数来反映切片中某物质含量。但滤光片的使用不仅会导致图像的颜色对比度降低，从而干扰医生后续诊断、增加误诊可能性，而且提高了支出成本。为了解决该问题，训练CGAN模型实现无滤光图到有滤光图的转换，取代硬件滤光片，并将训练好的模型称作数字滤光片。为提高数字滤光片性能，在生成器中添加全尺度跳跃连接以提高生成图像质量，在损失函数中添加积分光密度的约束以利于准确测量生成图像的细胞核积分光密度。实验表明，数字滤光片可达到硬件滤光片的滤光效果，生成图像在视觉效果上逼近真实滤光图像。在计算细胞核积分光密度的任务上，生成图像与均值为119的真实滤光图像偏差5.3%，表现效果分别比CycleGAN、pix2pix、pix2pixHD方法提升10.9%、3.2%、6.1%。
Abstract:Abstract:The pathology-assisted diagnosis system adopts the method of placing a filter on the slice to filter the light source and obtain the monochromatic parallel light, which is used to calculate the optical density parameter of the image to reflect the content of a certain substance in the slice.But the use of filter not only reduces the color contrast of the image, which interferes with the doctor’s follow-up diagnosis and increases the possibility of misdiagnosis, but also increases expenditure costs.In order to solve this problem, the CGAN model is trained to realize the conversion from unfiltered image to a filtered image, which is used to replace the hardware filter, and the trained model is called a digital filter.In order to improve the performance of the digital filter, a full-scale jump connection is added to the generator to improve the quality of the generated image, and the integral optical density constraint is added to the loss function to facilitate accurate measurement of the integrated optical density of the nucleus in the generated image.Experiments show that the digital filter can achieve the filtering effect of a hardware filter, and the generated image is close to the real filtered image in terms of visual effects.In the task of calculating the integral optical density of the cell nucleus, the generated image deviates 5.3% from the real filtered image with a mean value of 119, the performance is improved by 10.9%, 3.2%, and 6.1% compared with CycleGAN, pix2pix, and pix2pixHD methods.

PDF全文下载地址:

可免费Download/下载PDF全文