Performance of deep convolutional neural network approaches and human level in detecting mosquito species

Recently, mosquito-borne diseases have been a significant problem for public health worldwide. These diseases include dengue, ZIKA and malaria. Reducing disease spread stimulates researchers to develop automatic methods beyond traditional surveillance Well-known Deep Convolutional Neural Network, YOLO v3 algorithm, was applied to classify mosquito vector species and showed a high average accuracy of 97.7 per cent. While one-stage learning methods have provided impressive output in Aedes albopictus, Anopheles sinensis and Culex pipiens, the use of image annotation functions may help boost model capability in the identification of other low-sensitivity (< 60 per cent) mosquito images for Cu. tritaeniorhynchus and low-precision Ae. vexans (< 80 per cent). The optimal condition of the data increase (rotation, contrast and blurredness and Gaussian noise) was investigated within the limited amount of biological samples to increase the selected model efficiency. As a result, it produced a higher potential of 96.6 percent for sensitivity, 99.6 percent for specificity, 99.1 percent for accuracy, and 98.1 percent for precision. The ROC Curve Area (AUC) endorsed the ability of the model to differentiate between groups at a value of 0.985. Inter-and intra-rater heterogeneity between ground realities (entomological labeling) with the highest model was studied and compared to research by other independent entomologists. A substantial degree of near-perfect compatibility between the ground truth label and the proposed model (k = 0.950 ± 0.035) was examined in both examinations. In comparison, a high degree of consensus was assessed for entomologists with greater experience than 5-10 years (k = 0.875 ± 0.053 and 0.900 ± 0.048). The proposed YOLO v3 network algorithm has the largest capacity for support-devices used by entomological technicians during local area detection. In the future, introducing the appropriate network model based methods to find qualitative and quantitative information will help to make local workers work quicker. It may also assist in the preparation of strategies to help deter the transmission of arthropod-transmitted diseases.


60
Mosquito-borne diseases are a significant problem for human health. More than one million 61 cases have been recorded, with an estimated 400,000 deaths in 2018 [1]. These mosquito diseases 62 mostly included dengue fever, zika and malaria, which are prevalent in tropical and subtropical 63 areas. However, they can still spread and cause illness in other areas of the world due to human 64 mobility: globalization, labour movement and transport. In order to minimize morbidity and 65 disease-related mortality, the WHO has urged researchers to further develop more efficient 66 methods that can be used in the entomological field [2]. 67 Aedes genus mosquitoes (Ae. aegypti, Ae. albopictus, and Ae. vaxans) are pathogens with 68 the primary focus because they can spread human arboviral diseases: dengue, chikungunya, zika 69 and yellow fever [3,4]. In addition, Anopheles mosquitoes (An. dirus, An. minimus, An. sinensis 70 and An. maculatus) are vectors for human malaria parasites and can cause a high death rate [5,6]. 71 Culex mosquitoes can also spread all arboviruses; West Nile virus and human blood parasites  Automatic classification tools for mosquito species have been researched to better assist 86 local health personnel. Classification was successfully conducted using an analysis of image 87 characteristics and a flight tone for insects [17][18][19]  This study has been preceded by our previous work that identified gender and also species 120 of field-caught mosquito vectors [34]. The aim of this study is to choose the appropriate model

166
Of the total image set, the species of mosquitoes and non-mosquitoes is divided into 167 separate directories. These files have been randomly assigned to training/validation and testing 168 sets. They retained a raw test range of 10% for each folder. The remaining images were randomly 169 divided into training/validation or 90% for each folder, the same as before. These insect-specific To study how effective the chosen model was, it was then compared to human ability.

207
Inter-and intra-human variation in the identification of the tested image sets were assigned.

208
Inter and intra-rater variability in human-level 209 Inter-and intra-human variation was designed to allow observations between model and were completed with a two-month interval. In the first round, 15 independent raters with less than 213 5 years of experience were anonymously assigned to take the test. 10% % of the test image set 214 was selected and prepared for all raters via Google form (Fig 1). In addition, 25 independent raters 215 were recruited to take part in the second evaluation, which varied from the previous version CIs is measured using a non-parametric bootstrap approach of 1000-fold image re-sampling.

238
The degree of consensus between the model and the ground truth, as well as the 239 independent entomologists, was studied for exploratory observation by using two operational

251
Comparison of performance of the network model 252 The assessment of the two models were investigated based on a threshold probability, P(t), 253 from t 5% to t 95% . The predictions obtained from these models were provided by P class ≥ t the model capacity to distinguish between classes at a value of 0.985 (Fig 3). Hence, annotation  Inter-and intra-rater variability 302 Evaluation of performance was studied at the level of consensus of the ground truth 303 (entomological labeling); with the best model and connected to other independent entomologists.

304
Here, determine whether the model chosen was successful enough to be applied in a specific 305 environment. 10 % of the test datasets were randomly chosen and used for blind testing by all 306 human investigators; as inter-and intra-human variants. Approximately one month after the first 307 evaluation, 10 percent of the image re-selection from the test collection was circulated through 308 Google form to the same examiners for more intra-and inter-examination variability. have become more interested in automatic systems, although the ability of the system does not rely 331 on its processing time, but rather depends on data to learn the model within a time-limited frame.  The proposed YOLO v3 network algorithm provides great potential for rapid screening 374 and support devices for entomological technicians, especially during mosquito identification.

375
According to inter-and intra-human variability, the experiment has been accomplished by 376 encouraging them to analyze the image sample as being the same as the model test set. In the 377 future, qualitative and quantitative devices based on the best network model will help to make it 378 easier for local workers to perform quicker and to prepare strategies related to the advanced 379 prevention of mosquito-borne diseases. The data that support the findings of this study are available upon requested to the 394 corresponding author.