Logo image
Back
Promise and Impediments in the Use of Deep Learning for the Identification of Parasitic Lice
Journal article   Peer reviewed

Promise and Impediments in the Use of Deep Learning for the Identification of Parasitic Lice

Masoud A. Rostami, Stanislav Kolencik, Behnaz Balmaki, Daniel R. Gustafsson, Jason D. Weckstein, John Gausas, Ricardo L. Palma, Oldrich Sychra, Andrew D. Sweet, Kamila M. D. Kuabara, …
The Journal of parasitology, v 112(5), pp 547-556
25 Sep 2026
DOI: https://doi.org/10.1645/24-64

Abstract

Lice are common ectoparasites of both birds and mammals, including domesticated animals on which humans depend. Accurate identification of these parasites is necessary for those in public health, veterinary medicine, animal husbandry, ecology, and wildlife management. However, taxonomic expertise is increasingly in short supply. Deep learning (DL), a specialized subset of artificial intelligence, has been used successfully to aid nonspecialists in species-level identifications from digital images; however, unlike for many free-living organisms, there are relatively few images and taxonomically vetted collections of lice for DL model training. Available images are usually of slide-mounted specimens, with a growing number of images of unmounted specimens taken with a macro lens or using a dissecting microscope. Here, we use available imagery for a subset of bird lice from 12 genera commonly found on poultry that are often misidentified in the scientific literature. We use these images and computer vision approaches to test the potential of DL identifications for this group of ectoparasites and to identify the possibility for this approach with a group for which few images are available. We used a transfer learning method to leverage pre-trained models on large datasets and adapted these models to the specific task of louse detection with a focus on detailed feature analysis. We found the DenseNet-201 model performed the best on our data in terms of accuracy, efficiency, and robustness against overfitting, with a training accuracy of 99.2% and validation accuracy of 96%. Our 5-fold cross-validation confirmed this performance, with a macro F1 score of 0.95 and a weighted F1 score of 0.96. We also implemented an optimized training procedure, which contributed to generalization in the model. When we examined how well the model predicted the genus of previously misidentified images, we found that it correctly identified only 63.6% of the images, highlighting issues with generalization capabilities. We suggest reasons for inaccurate identifications in our study and encourage the louse taxonomic community to continue archiving high-quality taxonomically vetted images to increase the robustness of these models.

Metrics

1 Record Views

Details

Logo image