FecalNet: Automated detection of visible components in human feces using deep learning

• DOI: 10.1002/mp.14352
• Publication Date: 2020-06-25T02:58:02Z
• Authors: Qiaoliang Li; Shiyu Li; Xinyu Liu; Zhuoying He; Tao Wang; Ying Xu; Huimin Guan; Runmin Chen; Suwen Qi; Feng Wang
• Keywords: Identification; Feature (linguistics)

Purpose To automate the detection and identification of visible components in feces for early diagnosis gastrointestinal diseases, we propose FecalNet, a method using multiple deep neural networks. Methods FecalNet uses ResNet152 residual network to extract learn characteristics fecal microscopic images, acquire feature maps combination with pyramid network, apply full convolutional classify locate components, implement improved focal loss function reoptimize classification results. This allowed complete automation feces. Results We validated this database 1,122 patients. The results indicated mean average precision (mAP) 92.16% an recall (AR) 93.56%. (AP) AR erythrocyte, leukocyte, intestinal mucosal epithelial cells, hookworm eggs, ascarid whipworm eggs were 92.82% 93.38%, 93.99% 96.11%, 90.71% 92.41%, 89.95% 93.88%, 96.90% 91.21%, 88.61% 94.37%, respectively. times required by GPU CPU analyze image are approximately 0.14 1.02 s, Conclusion can It also provides framework detecting several other types cells clinical practice.