Case Segmentation of Nematode Microscopic Image Based on Improved Mask R-CNN

Abstract:
In order to obtain the experimental data of Caenorhabditis elegans, it is the most effective and intuitive means to observe and record the morphological changes of Caenorhabditis elegans. The traditional manual observation methods have large subjective errors, low efficiency and poor accuracy, so it is very important to realize efficient and automatic nematode morphological segmentation by computer. However, there are many impurities in the microscopic image of the nematode, which leads to the ambiguity of the main target. It is also difficult to accurately segment the nematode because the insect is transparent and the contrast between the insect and the surrounding environment is small. To solve these problems, a nematode micro image segmentation method based on improved Mask R-CNN is proposed. Firstly, the RoI Align layer is replaced by the PR RoI Pooling layer to further improve the accuracy of nematode target segmentation in the nematode micro image; After that, non maximum inhibition was improved to solve the problems of missed detection and false detection of nematodes. Compared with the Mask R-CNN before the improvement, the average accuracy (AP) is increased by 2.38% and the intersection union ratio (IoU) is increased by 12.66%. The experimental results show that the improved Mask R-CNN segmentation algorithm can significantly improve the accuracy of nematode micro images, and can accurately and effectively segment nematode targets in micro images.