Abstract:

Objective To develop an intelligent interpretation model for the results of indirect hemagglutination assay（IHA）of Schistosoma japonicum antibody in sera based on deep learning algorithms, and to preliminarily evaluate its efficacy so as to realize automatically and intelligently translating the results of schistosomiasis by diagnostic IHA. Methods Prepare and collect serum samples from rabbits (humans) with positive and negative results of Schistosoma japonicum, create IHA test result images with different levels of agglutination, and evaluate the agglutination results. The images with microtiter tray absent of IHA reaction were collected, and treated by enhancement, on which basis the image data sets were constituted. Then the intelligent interpretation model was created based on convolutional neural network PP-LCNet algorithms, and trained and tested respectively by training set and test set. The S. japonicum antibody gold-standard negative and positive rabbit sera diluted in different ratio (1∶5-1∶100) were detected by IHA, and the results were interpreted by three well-experienced professionals and the intelligent translating model independently. The performances of the model and professionals were evaluated as the specificity, sensitivity, accuracy, F1 score, diagnostic consistency (Kappa value). The receiver operation characteristic (ROC) curves were plotted to compare the coherence by the model and the professionals. Results In total, 15 956 images quantified by IHA were obtained, in which 2 271 were positive, 2 164 weak positive and 5 878 were negative reactions. Reaction was absent in 5 643 images of the microtiter trays. After the image enhancement, a total of 31 856 image sets were obtained, and divided training set (n=25 487) and test set (n=6 369). Efficacy evaluation showed that, except for the consistency in specificity (The agreement was 100.00% by the model and experts), the accuracy, sensitivity, F1 scores, Youden index and areas under ROC curves were 99.43%, 99.09%(97.84%, 100.34%), 99.54%, 0.990 9 and 0.995±0.009, respectively by the model results, which were slightly higher than the findings by professionals. The sensitivity of the intelligent interpretation model showed a statistically significant difference (χ²=6.125, 11.077, both P<0.05) between professional technicians A [95.45% (92.70%, 98.20%)] and B [93.18% (89.85%, 96.51%)], while there was no statistically significant difference (χ²=0.000, P>0.05) between professional technicians C [98.64% (95.31%, 97.17%)]. The results generated by the intelligent interpretation model highly agreed with gold standard. The consistency was somewhat better than the three professionals (Kappa= 0.988 vs. Kappa= 0.940, 0.910, 0.982, respectively, all P<0.05). Conclusion The intelligent interpretation model based on deep learning technology established in this study can accurately identify the results of S. japonicum antibody detection in sera by IHA, suggesting that it can be used for intelligently translating the IHA results for schistosomiasis japonica.

Key words: Schistosoma japonicum, Indirect Hemagglutination Assay, Deep learning, Image Recognition, Intelligent interpretation