Fig 1.
Representative digital images of ticks from the user-generated dataset compiled from iNaturalist, the Tick App, and the Wisconsin medical entomology laboratory.
The user-generated test set includes adult, nymphal and larval, male and female, fed and unfed tick specimens of the three most common human-biting ticks in the U.S.
Table 1.
Summary of images in the user-generated dataset.
Fig 2.
Cropping, resizing, and measuring the relative tick size (RTS) for user-generated images.
(a) Original resolution image of a female Dermacentor variabilis tick. (b) The cropped and resized image (224x224 pixels) with the relative tick size bounding box depicted in green (RTS = 67 pixels).
Fig 3.
Image augmentation procedure to increase the size of the user-generated training dataset.
Each image was re-cropped, flipped, and rotated to generate 15 visually distinct images. This process increased the size of the training dataset from 7,625 to 95,119 images.
Fig 4.
Representative images from the laboratory test set that includes high-quality images of tick specimens taken on a plain background.
The laboratory test set includes male and female (fed and unfed) specimens of the three most common human-biting ticks in the U.S.. The laboratory test set was sourced from an online database maintained by tickreport.com; a tick identification and testing service.
Table 2.
Best performing configuration results for four different network architectures.
Fig 5.
Application of confidence thresholds to TickIDNet predictions.
The classifier returns a probability for each tick species class with the highest probability prediction S(x) determining the class prediction C(x). If S(x) is lower than the computed species threshold τ, the image is not returned with the C(x) class prediction but is classified as “unsure” instead. These formulas are adapted from Villon et al. [39].
Table 3.
Class-specific performance statistics of TickIDNet evaluated on the user-generated test set (n = 1094) and the laboratory test set (n = 300).
Fig 6.
Post-processing of TickIDNet predictions on user-generated test set with species-specific confidence thresholds to increase the correct classification rate (CC) at the cost of more images being returned as “unsure.”
Three different sets of thresholds were calculated by applying 10%, 5%, 2% MC bounds to the validation set as described in Goal 2 of Villon et. al [39].
Fig 7.
Binary logistic regression of the relationship between a correct prediction and the relative tick size (RTS).
Fig 8.
Distribution of the relative tick size (RTS) of the user-generated test set.
Color represents correct (grey) and incorrect (black) prediction by TickIDNet. The total number of ticks identified was 1,094: 194 A. americanum, 625 D. variabilis, and 275 I. scapularis.
Fig 9.
The adjusted odds ratios for correct TickIDNet predictions based on tick characteristics and relative tick size (RTS).
Point estimate represents the adjusted odds ratio (aOR) and error bars the 95% confidence interval. The vertical line at aOR = 1 represents equal odds.
Fig 10.
Grad-CAM heatmap visualizations of TickIDNet predicted images show the model failing to assign importance to the relevant parts of the images when RTS is small.
The importance of shapes or regions in the image is ranked on a red-blue heatmap described by the scale on the right. (a) Dermacentor variabilis incorrectly identified as Amblyomma americanum with RTS = 15, (b) A. americanum incorrectly identified as Ixodes scapularis with RTS = 17, (c) I. scapularis correctly identified with RTS = 145, (d) I. scapularis correctly identified with RTS = 48.