Deep-Learning-Based Automated Morphology Analysis With Atomic Force Microscopy

Morphology analysis based on atomic force microscopy (AFM) imaging contributes to understanding the characteristics of specimens more deeply. The preliminary and crucial step of implementing morphology analysis is to precisely segment the target area from the complex background. In this study, an automated AFM image segmentation strategy based on a well-designed U-shaped neural network is proposed to achieve accurate and robust segmentation for AFM images of different samples, thus realizing morphology analysis in micro-nano scale. Specifically, the centralized information interaction strategy cooperated with a two-path attention module is introduced to realize efficient cross-scale information interaction, which can fundamentally avoid the negative effects induced by spatial interpolation. Besides, the global information flows are adopted to guide the global information extracted by atrous spatial pyramid pooling to each level of the top-down pathway, which ensures that the high-level semantic information is not diluted during the top-down transmission process, thus locating the target area more precisely. Moreover, an AFM image dataset is constructed to train the network, which will be available online for free to facilitate other data-based AFM research. The segmentation results demonstrate that the proposed strategy has better performance on multiple AFM images compared with traditional Otsu method, fully convolutional network and U-Net. The application of the proposed method is carried out to exhibit the effectiveness in automated morphology analysis. Note to Practitioners-Despite the growing demand of AFM-based morphology analysis in many fields, the automated analysis is still lacking limited by accuracy and robustness of AFM image segmentation. Since manual segmentation, sometimes tedious and time-consuming, heavily depends on the personal judgment, it is thus necessary to develop automated segmentation methods. Although traditional automated segmentation algorithms have good performance on certain types of images, they may be difficult to apply in different scenarios, especially for micro-nano images, due to the limited robustness. Therefore, this paper proposes an automated image segmentation algorithm based on an improved U-shaped neural network to achieve accurate morphology analysis for AFM images. The proposed automated morphology analysis workflow will be a practical tool to help reduce human workload and subjective errors, as well as enhancing the accuracy and robustness of the analysis process. In addition, the constructed AFM image dataset can greatly facilitate the research on data-based AFM image analysis of other practitioners. Moreover, practitioners can benefit from our algorithm to improve the accuracy and robustness of image segmentation in other practical applications.