1. Recognition of neonatal biliary atresia based on transformer.
- Author
-
QIN Zhong-Han, AI Cheng-Bo, TAN Chao-Qun, LIU Hong, DU Wen-Chao, YANG Hong-Yu, WU Zhi-Hong, and CHEN Hu
- Subjects
BILIARY atresia ,CONVOLUTIONAL neural networks ,GALLSTONES - Abstract
Neonatal biliary atresia is one of the most common fatal diseases in neonates, with higher incidence rates in Asia than in other parts of the world. Early detection and treatment of neonatal biliary atresia are crucial, yet the lack of professional pediatricians and auxiliary diagnostic and treatment methods can cause parents to miss the best treatment window. To address this issue, this paper develops a predictive algorithm with practical application value that uses neonatal stool pictures to predict whether the newborn has neonatal biliary atresia and reminds parents to visit a doctor in time. To achieve higher recognition rates in practical scenarios, the algorithm in this paper is developed based on a real-scene dataset of newborn fecal images. First, we designed a self-attention network model BANet (Biliary Atresia Network), which will combine shallow features and deep features of pictures to get better classification. To address issues as dark or overexposed images, we developed an automatic brightness adjust-ment algorithm by analyzing the brightness distribution of the dataset. Furthermore, we added a shadow data enhancement method duiring training to mitigate the inference of shadows on recognition results. In order to verify the effectiveness of the algorithm proposed in this paper, we design a comparison test with doctors. The results show that BANet out performed doctors in objective evaluation indicators such as the recognition rate of four classifications, the recognition rate of two classifications, specificity and sensitivity. The proposed BANet can effectively use the color, abnormal points and other information in the picture, by compensating the brightness of the picture, the accuracy and robustness of the overall algorithm are improved and good results have been achieved in practical application scenarios. [ABSTRACT FROM AUTHOR]
- Published
- 2023
- Full Text
- View/download PDF