Improved segmentation of collagen second harmonic generation images with a deep learning convolutional neural network

Collagen fibers play an important role in both the structure and function of various tissues in the human body. Visualization and quantitative measurements of collagen fibers are possible through imaging modalities such as second harmonic generation (SHG), but accurate segmentation of collagen fibers is difficult for datasets involving variable imaging depths due to the effects of scattering and absorption. Therefore, an objective approach to segmentation is needed for datasets with images of variable SHG intensity. In this study, a U-Net convolutional neural network (CNN) was trained to accurately segment collagen-positive pixels throughout SHG z-stacks. CNN performance was benchmarked against other common thresholding techniques, and was found to outperform intensity-based segmentation algorithms within an independent dataset, particularly at deeper imaging depths. These results indicate that a trained CNN can accurately segment collagen-positive pixels within a wide range of imaging depths, which is useful for quantitative SHG imaging in thick tissues.

Automated summary

Collagen fibers play a vital role in various tissues, but accurate segmentation is challenging due to variable imaging depths. This study employed a U-Net convolutional neural network for precise segmentation of collagen fibers and demonstrated superior performance compared to other thresholding techniques, particularly at deeper imaging depths.