Speaker
Description
"Introduction: Over the last decade, an increasing relevance has been given to the use of semi-quantitative and quantitative methods to evaluate histological and immunohistochemical images through image analysis software. Most of this software only require image input. However, the usability of classic image processing, including classic segmentation algorithms, is limited. Machine learning could provide a new tissue engineering tool that may allow a feasible, objective, and automated approach that serve as quality control of bioengineered tissues. In this context, the aim of this study is to develop a supervised machine learning approach for the accurate and automatic quantification of collagen fibers in human oral mucosa substitutes generated by tissue engineering and stained with Picrosirius red for collagen detection.
Methods: Human oral mucosa substitutes were generated by tissue engineering. First, oral mucosa keratinocytes and fibroblasts cell cultures were established from small oral mucosa biopsies. Then, a stromal substitute was fabricated using 0.1% fibrin-agarose and epithelial cells were cultured on top1. Native oral mucosa samples were used as control, and both types of samples were stained with Picrosirius red. Then, a neural network was developed and trained from scratch using 10 large histological images. During training, we took random patches and applied different types of data augmentation to them. Following this process, we were able to create in real time an uncountable amount of unique small patches from the original images. Once trained, we tested the network in a test set of histological images. To obtain the mask of a specific area, sub-masks were obtained and finally combined to obtain the entire area.
Results: Application of the automated quantification system developed here allowed us to accurately identify the target structures in each histological image. Under leave-one-out cross-validation over the histological images, the method yielded over 90% pixel accuracy and collagen precision in training and over 85% in pixel accuracy and collagen precision in set.
Conclusions: The proposed machine learning model based in neural networks was able to segment collagen in histological images based on semantic information, instead of the more classic color segmentation used in the field. The developed models demonstrate to be more robust to color outliers and is able to produce a better segmentation. Furthermore, the segmentation produced by the network is fully automatic, which also reduces the tedious process of having to manually fine-tune a color range to get a correct segmentation. This quantitative and automated approach open a new window on quality control techniques for bioengineered human oral mucosa tissues."
52354516564
