基于深度学习的数字病理扫描系统单次曝光自动对焦方法

Whole slide imaging via one-shot deep autofocusing

全切片数字成像(whole slide imaging, WSI)是病理切片数字化的核心技术,其自动对焦的速度和精度决定了WSI系统的性能.然而,传统的自动对焦方法需要拍摄多张离焦子图像创建图像堆栈,或者需要复杂的硬件调制光学系统,从而限制了WSI在实际场景中的应用.本文设计了基于深度学习的数字病理扫描系统单次曝光自动对焦方法,对子图像逐个进行网络虚拟自动对焦,将单张离焦子图像通过网络直接生成准焦子图像.本方法仅需要在任意离焦距离下的单次曝光拍照即可,从原理上避免了重复的对焦运动和相机曝光过程.实验结果表明,本方法具有高通量、高速度、低成本、实用性强、可线下处理等优点.

Whole slide imaging(WSI) is the core technology of pathology digitization, and the performance of WSI is determined by the speed and accuracy of autofocusing. However, the conventional autofocusing methods need to shoot multiple defocus tiles to create an image stack, or require a complex hardware modulation optical system, thereby restricting the applications in the real scenario of WSI. In this paper, we develop a deep one-shot autofocusing network of WSI for tile-wise virtual autofocusing to generate the in-focus image directly from one possible defocus image. Our method works with one image taken at any defocus distance, avoiding repeated focusing movements and camera shooting processes in principle. Experimental results demonstrate that our scheme enjoys the merits of high throughput, high speed, low cost, compatibility and offline processing.