Application of deep learning and threshold segmentation in nematode recognition

Objective To explore whether deep learning and threshold segmentation based methods can improve the efficiency and accuracy of behavioral experiments to realize the statistics of the number and aggregation degree of Caenorhabditis elegans in images, and facilitate the high-throughput screening of potential drugs for autism using nematodes and other small model organisms. Methods A total of 5000 samples of culture media containing single nematodes were collected. Then nematode recognition and detection algorithms were designed through manual counting, and deep learning and threshold segmentation method, respectively, so as to judge the count and aggregation degree of nematodes in culture media. Meanwhile, analysis time and accurate rate were compared. Results Based on deep learning and threshold segmentation algorithm, we designed a software that automatically calculated the quantity of nematodes and aggregation degree. The software was able to detect each image within milliseconds, with a remarkably increased detection rate, compared with 100 seconds or more spent for artificial statistics. The software also exhibited good detection accuracy, with over 90% accuracy for single worm detection. Conclusions Compared with manual counting, the software based on deep learning and threshold segmentation algorithm can significantly improve the efficiency of Caenorhabditis elegans counting and aggregation degree judgment. The algorithm based on deep learning and threshold segmentation in this study will significantly improve the screening efficiency of neurobehavioral drugs related to autism.