Abstract
Dog breed is fundamental health information, especially in the context of breed-linked diseases. The standardization of breed terminology across health records is necessary to leverage the big data revolution for veterinary research. Breed can also inform clinical decision making. However, client-reported breeds vary in their reliability depending on how breed was determined. Surprisingly, research in computer science reports that AI can assign breed to dogs with over 90% accuracy from a photograph. Here, we explore the extent to which current research in AI is relevant to breed assignment or validation in veterinary contexts. This review provides a primer on approaches used in dog breed identification and the datasets used to train models to identify breed. Closely examining these datasets reveals that AI research uses unreliable definitions of breed and therefore does not currently generate predictions relevant in veterinary contexts. We identify issues with the curation of the datasets used to develop these models, which are also likely to depress model performance as evaluated within the field of AI. Therefore, expert curation of datasets that can be used alongside existing algorithms is likely to improve research on this topic in both fields. Such advances will only be possible through collaboration between veterinary experts and computer scientists.
Breed is routinely collected in the course of veterinary care. It is especially relevant in canine health due to associations between breed ancestry and genetic disease in both purebred and mixed-breed dogs.1–5 Breed has long informed clinical decision making1,6,7; for example, breed-specific screening for genetic diseases has become a common practice in dog breeding and preventative care. In the age of big data, with large-scale analyses of electronic health records (EHRs) seeking to discover breed-trait associations, accurate entry of breed data is increasingly important.8 Because terminology used to describe breeds varies widely,9 breed data standardization is a target of work in veterinary informatics.9–11
However, harmonization is not the only challenge to standardizing breed data. Different approaches to determine breed can yield different labels, with some more reliable than others. Possible frameworks include pedigree, genetic testing, and visual identification. Variation in how breed is determined introduces challenges beyond the harmonization of terminology for efforts to standardize breed labels in EHRs.
The most reliable source of breed is pedigree. In pedigreed dogs, the lineage is recognized by a breed organization. However, definitions vary; for example, the American Kennel Club (AKC) recognizes a different set of breeds and uses different terminology for others than does the Fédération Cynologique Internationale.9,12 Breed organizations define their own breed standards, and criteria differ. Additionally, public perceptions of breed can also differ from official definitions: for example, a “mini-goldendoodle” is formally a mixed-breed dog, and dogs without pedigrees may still be described as “purebred”12 (eg, backyard-bred dogs). Therefore, reconciling reported breed with formal definitions is complicated, even for purported purebreds.
Genetic analysis provides an alternative perspective on breed. Because breeds are, to varying extents, genetically siloed,13 genetic variants present in a dog's DNA point to its ancestral breed(s).14–18 Direct-to-consumer (D2C) genetic tests of breed ancestry are widely available, but products yield different results, even for pedigreed dogs.19 Direct-to-consumer ancestry tests differ from targeted genetic tests used in veterinary practice (eg, Neogen): they focus on estimating breed composition (though some do also evaluate genetic health), whereas clinical genetic tests target specific disease variants related to health.8,20,21 Also, unlike with clinical genetic tests, many veterinarians are wary of interpreting D2C results.20
Finally, many dogs are assigned a breed based on appearance.22–24 This approach is flawed: studies22,23,25,26 report low concordance between breed(s) assigned visually and genetic ancestry. Canine welfare organizations therefore recommend against visual assignment of breed,1,27 but the practice remains common.22
Therefore, standardizing breed in EHRs requires not only harmonizing the terminology used but also assessing the source and reliability of breed information. While D2C breed ancestry testing offers a potential strategy to validate breed data, the high cost and the long turnaround time render this approach suboptimal, especially for veterinary EHR studies analyzing thousands to millions of records. Additionally, the lack of concordance among different tests19 and clinician mistrust of the industry20 means that, at present, tests do not fill this role.
In contrast, research in computer vision, an area of AI focused on extracting information from visual media, such as images or videos, frequently reports very high accuracy (> 90%) in assigning breed to dogs based on appearance alone. Because selective breeding and thus genetic ancestry influence both breed specific health risks and physical traits,28,29 health and appearance are also likely to be correlated. Artificial intelligence, which produces results in seconds, would offer a much more viable approach for validating breed data in EHRs at scale. Such a tool could also support veterinarians making real-time decisions where breed is important (eg, screening for genetic diseases).
Here, we explore whether AI has actually solved the problem of visual dog breed identification (DBI) as suggested by the literature. First, we provide a primer on computer vision and, specifically, the DBI problem. Next, we critically evaluate results from DBI studies, with specific attention to the image datasets used to train and evaluate the models. Finally, we propose that despite significant potential in this area of research, it currently suffers from a lack of input from domain experts, like veterinarians. In the current era, AI is becoming increasingly prevalent in all areas. Just as veterinarians are often asked to interpret the results of D2C genetic tests they did not order,20 understanding AI's perspective on breed assignment will not only facilitate discussions of such results with patients but, we hope, advance the field of veterinary health AI.
What Can AI Offer DBI?
With today's computing power, it is increasingly feasible for AI models to extract patterns hidden to the human eye. Though veterinary medicine is just beginning to benefit from AI, it is used to facilitate the discovery of patterns in human disease, the identification of candidate treatments, and more.30,31 One of the major concerns with using AI in medicine is that it mines patterns from whatever data it is provided, so biases in the data result in biased models. As an example, Amazon developed an AI-based tool to help with hiring decisions and trained it using prior applications and the associated hiring decision.32 Even though they removed demographic information, because AI is such a powerful pattern detector, it picked up on an implicit pattern: women were hired less often. The AI began screening out applicants whose resumes mentioned women-focused activities (eg, the Society of Women Engineers) or women's colleges.32 This concerning reproduction of societal biases reflects the importance of careful data curation when training AI models.
Bias is not a problem unique to AI. Errors in human judgment have been widely studied in psychology and are often understood through heuristics, which are logical shortcuts that make it easier to interpret complex information.33 One example is the representativeness heuristic, related to stereotyping, where people make judgments biased by similarity to a prototype.34 The representativeness heuristic likely explains some errors people make in visual breed identification. For example, 24% of animal experts surveyed labeled a white shelter dog with black spots as a purebred Dalmatian.25 Dalmatians are notable for their spots, but purebred Dalmatians are rare35; therefore, it is much more likely that a dog in the shelter with this coat color is a mixed-breed dog as ultimately supported by DNA analysis.25
Dog breed identification is therefore a candidate problem where AI could provide benefits. In human studies,34 heuristic-based judgments are contrasted with analytical evaluations, such as labeling a spotted dog as mixed breed because Dalmatians are rare. When designing AI models, appropriate selection of data and specification of models can mitigate the influence of these types of logical fallacies.36 Thus, a model trained on diverse examples of spotted dogs would not be expected to preferentially associate black spots with Dalmatians. Therefore, to ensure the accurate development and interpretation of AI models, it is necessary to understand how the training process shapes a model.
Basics of AI in DBI
Computer vision research in DBI typically uses a type of AI called convolutional neural networks (CNNs). Neural networks are inspired by neuroscience.37 Just as neurons pass signals via synapses, neural networks have connected layers. A layer is composed of neurons that receive and process input before passing it along, like an action potential traverses a neuron from its axon to its dendrite. Neural networks even have an activation function that determines whether or not a signal fires from 1 layer (neuron) to the next. This architecture allows neural networks to learn complex patterns in data.
Convolutional neural networks are initially built using vast, general-purpose datasets representing many concepts. These datasets contain 2 components: images and labels. The label can be thought of as the answer to the question “what does this image show?” As a CNN trains, it makes a prediction about what each image shows. This prediction is then compared to the associated label. The model is updated based on the accuracy of the prediction. This process repeats many times with different image/label pairs. Over time, individual layers within a CNN learn specific patterns in data, with earlier layers capturing very general features, like colors or edges, and later layers combining the outputs of earlier layers to construct more specialized features. Thus, predictions improve as the model learns to recognize features and patterns of features associated with different labels.
Such models, called pretrained models, have learned to recognize patterns in a very general dataset. To adapt a model to a specific task, it must be fine-tuned using additional labeled images from the domain of interest.38 Receiving feedback only on a specific problem allows the model to adjust the weights to specifically improve performance on, for example, DBI. Fine-tuning thus produces a version of the pretrained model optimized to a specific task.
After a model is trained, its ability to predict the contents of new images is evaluated. The test images are also labeled, but the labels are hidden from the model. Instead, the model's predictions are compared against the labels to evaluate performance. Two common metrics of performance are accuracy, or the proportion of correct answers, and loss, which is a fine-grained scoring system used during the training stage.38 The testing phase provides insight into how well the models achieves the ultimate goal: to reliably classify data when labels are unknown.
Why does computer vision study dogs?
Dog breed identification is a classic problem for computer vision that emerged when dogs were discovered to be a particularly challenging category of objects to identify in images.39,40 Identifying dog breeds is an example of fine-grained image classification (FGIC), where a category (dog) can take on many distinct patterns (breeds). Fine-grained image classification is especially difficult when differences between classes are small or variation within categories is large.41,42 Both are typical of DBI: as part of a single species, breeds share many similarities, and dogs within a breed can have different ages, coats, sizes, grooming, etc.43 These features made DBI a perfect challenge for studying how algorithms can be used to solve FGIC.
In the past, research in FGIC had to carefully select what information to include in models because image processing is so computationally expensive. Modern computing power makes it feasible to build very large networks with a large number of layers and parameters, allowing models to be tuned very precisely during training on huge datasets. Current research in DBI reports accuracies over 90%.44–48 Research today focuses on comparing effects of different CNN architecture (layouts of layers) on performance, with marginal gains (eg, up to 92.4%44) achieved by designing complex, hybrid architectures. However, as described above, the data used for pretraining and fine-tuning is also an important part of a model's performance. Therefore, it is necessary to examine the datasets these models are trained on to fully understand the results reported by these studies.
Benchmarking datasets for DBI
Dog breed identification is 1 of several problems that computer vision uses to study AI for FGIC. Dog breed identification studies today typically use pretrained models that are also used for other problems in computer vision. Many of these pretrained models are trained on ImageNet, a massive corpus containing hundreds of images for each of 80,000 concepts.49 To build ImageNet, researchers queried image search engines with terms from WordNet, a lexical hierarchy of concepts developed by AI researchers.49,50 These images were in turn evaluated for relevance by participants in Amazon's Mechanical Turk (MTurk; https://www.mturk.com), a marketplace that crowdsources human judgment by paying people to complete tasks.49 Dogs and cats, as well as specific breeds, were among the concepts rated. High rates of inter-rater disagreement indicated that reviewers struggled to differentiate animals at the breed level.49 Ultimately, researchers chose breed labels based on majority opinion.49 These image/label pairs are now fundamental to research in computer vision.
In machine learning, datasets used as standard representations of a problem are called benchmarks. Benchmarks standardize a task to allow direct comparison of different algorithms’ performances. For FGIC, a benchmark dataset is typically used to fine-tune and then evaluate performance by 1 or more pretrained models. Dog breed identification uses 3 primary benchmarks that were developed at different times and include different numbers of images and breeds. One thing that all 3 have in common is that they contain image/label pairs taken from ImageNet.
The original DBI benchmark, Stanford Dogs,43,51 contains exclusively images from ImageNet. Researchers selected all subclasses of the Canis familiaris class that did not have any subclasses of their own and then removed any likely to include images of multiple breeds, like “puppy.”43 Each image was checked “to confirm whether or not it matched images from Wikipedia and shared similar features to the other images in the same category.”43 Ultimately, 22,000 images representing 120 breeds (150 to 200 images/breed)43 were retained (Figure 1).
Image frequency across datasets by breed clade. Each bar represents a breed, and colors match those used in the published breed phylogeny. Clades that were not named54 are described by their position within the phylogeny (OtherAsian, EurAsian). Some groups (outgroup, basal, unknown) represent breeds/species that were not included in the phylogeny. Basal = Village dog “breeds.” Outgroup = Nondog species. Unknown = Breed not analyzed.54
Citation: American Journal of Veterinary Research 86, S1; 10.2460/ajvr.24.10.0315
The other benchmarks curated additional images to complement those in ImageNet. For the Columbia Dogs Dataset, images were retrieved from the photo-sharing site Flickr and Google image search.52 Breed labels were determined by majority opinion among “multiple” MTurk workers.52 The final dataset included 8,351 images across 133 AKC-recognized breeds52 (< 100 images/breed; Figure 1). For Tsinghua University's 2020 dataset, images were collected with a “data capture system” involving user-uploaded photos from 3 Chinese cities and from image searches and the social-networking site Baidu.46 Images were assigned a breed label via “expert review.”46 The curators sought to represent Chinese dog breed diversity both in terms of the 130 breeds included and the relative representation of breeds, with frequencies ranging from 200 to nearly 7,500.46 The final dataset includes over 70,000 images.46
In order to compare the breeds contained in these 3 datasets, we used the same approach as described previously19 to harmonize breed labels across datasets using 2 sources: the Vertebrate Breed Ontology (VBO), an expert-curated family tree of agricultural and companion animals that standardizes conceptually identical breed names,9,53 and a dog breed phylogeny indicating genetic relationships among breeds.54 This approach elucidated the numbers of images included for different breed or clade categories across the 3 datasets (Figure 1).
Limitations of data curation
The most significant limitation of these benchmarks is the data curation. For all 3, labels were assigned and/or validated visually, usually by nonexperts. Because people, including experts, cannot reliably predict breed based on appearance,22,23,25,26 the labels assigned to images in these datasets are likely to contain errors. As a result, computer vision studies are training models to predict what a human would think a dog's breed is rather than the dog's actual breed ancestry. Another issue is that these datasets are built with images of unknown provenance; a dog being labeled as a specific breed on Flickr does not mean that it is pedigreed or provide information about which breed organization(s)'s definition of the breed is being used. Furthermore, because breed definitions differ among breed groups and over time, some dogs might adhere to some breed organizations’ standards but not others, a challenge that is especially notable given the global nature of MTurk's workforce. Therefore, despite reports of accuracy over 90%, it is unlikely existing models for DBI have been trained to make predictions relevant in a veterinary context.
Digging deeper into DBI
Computer vision and veterinary medicine approach DBI very differently. For AI, dog breeds present a challenging FGIC problem: small differences between classes and wide variation within classes. On the biomedical side, breed is an important piece of health data, and standardizing collection of this data would facilitate the identification and application of breed-trait associations.8 Breed predictions made by current models from DBI are not likely to be relevant to veterinary medicine. However, computer vision has reported great success at visual assignment of breed, albeit using an unreliable definition of breed. Therefore, the approaches developed for DBI by computer vision could be useful to efforts to validate breed data in EHRs with better data for fine-tuning. In particular, we suggest 2 ways to improve dataset curation that would likely increase both model performance and the relevance of models to animal health.
Validating labels
Currently, the main issue with the benchmarking datasets is that the images were labeled using methods known to be unreliable (namely, human judgment). When a model guesses that an image shows a breed that does not match the label, it is currently impossible to know which is incorrect. Using specific, externally valid criteria to define breed would ensure that models are trained and tested on the actual question of interest. One option would be to train and test models on images of purebred dogs registered with a specific breed organization. This approach would eliminate noise introduced by different breed organizations having different breed standards. However, a classifier trained to identify, for example, only AKC-registered dogs would have limited external utility as most dogs are not expected to conform to breed standards.
Another option is to use genetic ancestry results. Genetic ancestry testing allows unpedigreed and mixed-breed dogs to be assigned breed labels. Fine-tuning using images of dogs with known genetic ancestry would allow models’ predictions to be validated using an external ground truth. Using images of dogs with single-breed genetic ancestry would provide the same advantages as using images of pedigreed dogs while also including dogs outside any specific breed standard (eg, backyard-bred dogs or registered dogs from other countries). Thus, incorporating genetic ancestry results into the curation of datasets for DBI is expected to both remove noise from the training data and aid in this research addressing questions relevant to the needs of veterinary medicine.
Including only dog breeds
Beyond labels, noise is also introduced to existing benchmarking datasets by the inclusion of classes that skew the distribution of diversity across classes. The Stanford and Tsinghua datasets both include classes of dogs, such as free-living colonies of village dogs (eg, Chinese Rural Dogs or Canadian Eskimo Dogs) that have never been bred to a standard and are more genetically diverse than dog breeds (Figure 1).55 These datasets also include nondog species: the dhole (Cuon alpinus), which diverged from the dog lineage over 5 million years ago,56 and the African wild dog (Lycaon pictus), separated for 2 to 3 million years.57,58 Another outgroup in these datasets is the dingo, which is sometimes considered a separate species (C dingo) or a subspecies of wolf (C lupus dingo); regardless of taxonomy, dingoes are genetically distinct from modern dogs breeds.59,60 Animal in these outbred classes, which are not bred to a standard, are expected to show a broader range of apperances than in dog breeds. Because higher within-class diversity makes FGIC more challenging, including these classes complicates the problem without any obvious benefit. Thus, the inclusion of these classes works against the goals of both fields.
The Tsinghua dataset, in particular, also amplifies the issue of between-class similarity. In 2 cases, this dataset uses multiple class labels to describe a single breed. The terms “German Shepherd” and “Black Sable” describe the same breed within the VBO, mapping onto the VBO identifier VBO:0200577, whereas “Miniature Poodle” and “Teddy” both describe VBO:0201051. Black Sable describes a German Shepherd coat color, and “Teddy” refers to a grooming style used in China for Miniature Poodles with specific fur color and muzzle shapes.61,62 Differentiating categories below the level of breed is exceptionally challenging, and subdividing a breed increases the complexity of the patterns the model must learn with no apparent benefit. Thus, including only valid breeds under specific, dog diversity–informed criteria (eg, AKC breeds) would be expected to improve model performance while retaining all of the classes relevant to applications in veterinary medicine.
The future of DBI
Inaccurate labels are a major source of error and bias in machine learning, so much so that a common saying is “garbage in, garbage out.”63 As discussed previously, AI perpetuates biases in the training data unless these effects are deliberately mitigated.63,64 A model cannot produce predictions better than the data it is trained on, so improperly curated breed data is a major concern for DBI. With AI used widely in criminal justice65 and government surveillance,66,67 it is not difficult to imagine AI being used to identify prohibited breeds or crossbreeds of dogs. Just as with D2C genetic tests that assign breed incorrectly,19 AI models have the potential to make these ethical issues even more fraught. To this point, some researchers describe concerning motivations for work in DBI that perpetuate misconceptions about relationships between breed, appearance, and behavior. One such paper argues, in the context of breed identification, that if “dangerous pets can be distinguished in time, it can bring more security to people and avoid some people being bitten.”68 In contrast, genetic studies do not support breed stereotypes as breed explains only 9% of inter-individual behavioral variation.16 In such contexts, models incorrectly assigning of dog breeds potentially introduces concerns far beyond EHR standards.1,24 Therefore, involving veterinarians and other animal experts in research in this area is necessary not only to ensure tools hold relevance outside of computer vision, but it is also critical in preventing the development of models that perpetuate real harm.
Conclusions
For over a decade, computer vision has studied how to identify dog breeds based on appearance and today reports over 90% accuracy on this task. At face value, it seems that AI has developed models that could be adopted to validate breed in veterinary record keeping. However, close review of this area of research reveals that the datasets used for DBI were curated without input from animal experts. Thus, despite years of work on a problem relevant to veterinary research, and despite computer vision's reported success on DBI, this work does not address the needs of veterinary medicine. However, the algorithms developed for FGIC perform well even when trained on noisy datasets, suggesting that informative models might be possible with the right training data. These same sources of noise are also likely to artificially deflate performance as defined in computer vision. As a result, integrating the perspectives of computer vision and breed diversity yields several possible avenues that could yield beneficial results across both fields. As veterinary research begins to embrace the big data revolution, greater collaboration is needed between veterinary experts and AI researchers to ensure that we are asking the right questions and building models that address the most important challenges in animal health and welfare. Visual assignment of dog breed is one area of AI perfectly suited to this type of crossdisciplinary innovation.
Acknowledgments
The authors thank Tana Liu and Elizabeth DelMonico for work mapping breed labels onto Vertebrate Breed Ontology terms and Xiaoyi Wu for summarizing convolutional neural network architectures.
Disclosures
The authors have nothing to disclose.
ChatGPT was used to make minor modifications to text and images. Specifically, in writing an early draft, the authors used it to workshop high-level explanations of technical machine learning concepts, asking it to evaluate phrases for accuracy and accessibility. Its feedback was used to iteratively edit these explanations at the level of phrase or sentence. Similarly, team members wrote the initial code used to generate Figure 1, then asked ChatGPT to tweak the code to achieve the desired esthetics (eg, “Modify this Matplotlib code to remove tick marks from the y-axis”).
Funding
Student travel to present an early version of this work at the Symposium on Artificial Intelligence in Veterinary Medicine was supported by the Schultz Foundation of Smith College and SAVY conference organizers (to Glenvelis Perez).
ORCID
G. Perez https://orcid.org/0009-0000-6092-1547
Y. He https://orcid.org/0009-0007-5692-8608
Z. Lyu https://orcid.org/0009-0002-5325-6838
Y. Chen https://orcid.org/0009-0009-2640-8235
H. M. Rando https://orcid.org/0000-0001-7688-1770
References
- 1.↑
Simpson RJ, Simpson KJ, VanKavage L. Rethinking dog breed identification in veterinary practice. J Am Vet Med Assoc. 2012;241(9):1163–1166. doi:10.2460/javma.241.9.1163
- 2.
Parker HG. Genomic analyses of modern dog breeds. Mamm Genome. 2012;23(1–2):19–27. doi:10.1007/s00335-011-9387-6
- 3.
Zierath S, Hughes AM, Fretwell N, Dibley M, Ekenstedt KJ. Frequency of five disease-causing genetic mutations in a large mixed-breed dog population (2011–2012). PLoS One. 2017;12(11):e0188543.
- 4.
Hayward JJ, Castelhano MG, Oliveira KC, et al. Complex disease and phenotype mapping in the domestic dog. Nat Commun. 2016;7:10460. doi:10.1038/ncomms10460
- 5.↑
Donner J, Freyer J, Davison S, et al. Genetic prevalence and clinical relevance of canine Mendelian disease variants in over one million dogs. PLoS Genet. 2023;19(2):e1010651. doi:10.1371/journal.pgen.1010651
- 6.↑
Paynter AN, Dunbar MD, Creevy KE, Ruple A. Veterinary big data: when data goes to the dogs. Animals (Basel). 2021;11(7):1872. doi:10.3390/ani11071872
- 8.↑
Lustgarten JL, Zehnder A, Shipman W, Gancher E, Webb TL. Veterinary informatics: forging the future between veterinary medicine, human medicine, and One Health initiatives—a joint paper by the Association for Veterinary Informatics (AVI) and the CTSA One Health Alliance (COHA). JAMIA Open. 2020;3(2):306–317. doi:10.1093/jamiaopen/ooaa005
- 9.↑
Toro S, Matentzoglu N, Mullen KR, et al. Classifying animal breeds with the vertebrate breed ontology (VBO). In: Proceedings of the International Conference on Biomedical Ontology. International Conference on Biomedical Ontology 2022. Accessed December 30, 2024. https://icbo-conference.github.io/icbo2022/papers/ICBO-2022_paper_1882.pdf
- 10.
Santamaria SL, Zimmerman KL. Uses of informatics to solve real world problems in veterinary medicine. J Vet Med Educ. 2011;38(2):103–109. doi:10.3138/jvme.38.2.103
- 11.↑
O’Neill DG, Church DB, McGreevy PD, Thomson PC, Brodbelt DC. Approaches to canine health surveillance. Canine Genet Epidemiol. 2014;1:2. doi:10.1186/2052-6687-1-2
- 12.↑
Fossati P, Ruffo G. Purebred dogs and cats: a proposal for a better protection. J Vet Behav. 2021;45:44–50. doi:10.1016/j.jveb.2021.05.009
- 13.↑
Lampi S, Donner J, Anderson H, Pohjoismaki J. Variation in breeding practices and geographic isolation drive subpopulation differentiation, contributing to the loss of genetic diversity within dog breed lineages. Canine Genet Epidemiol. 2020;7:5. doi:10.1186/s40575-020-00085-9
- 14.↑
Dreger DL, Rimbault M, Davis BW, Bhatnagar A, Parker HG, Ostrander EA. Whole-genome sequence, SNP chips and pedigree structure: building demographic profiles in domestic dog breeds to optimize genetic-trait mapping. Dis Model Mech. 2016;9(12):1445–1460.
- 15.
Vonholdt BM, Pollinger JP, Lohmueller KE, et al. Genome-wide SNP and haplotype analyses reveal a rich history underlying dog domestication. Nature. 2010;464(7290):898–902. doi:10.1038/nature08837
- 16.↑
Morrill K, Hekman J, Li X, et al. Ancestry-inclusive dog genomics challenges popular breed stereotypes. Science. 2022;376(6592):eabk0639. doi:10.1126/science.abk0639
- 17.
Parker HG, Kim LV, Sutter NB, et al. Genetic structure of the purebred domestic dog. Science. 2004;304(5674):1160–1164. doi:10.1126/science.1097406
- 18.↑
Leroy G, Verrier E, Meriaux JC, Rognon X. Genetic diversity of dog breeds: between-breed diversity, breed assignation and conservation approaches. Anim Genet. 2009;40(3):333–343. doi:10.1111/j.1365-2052.2008.01843.x
- 19.↑
Rando HM, Graim K, Hampikian G, Greene CS. Many direct-to-consumer canine genetic tests can identify the breed of purebred dogs. J Am Vet Med Assoc. 2024;262(5):1–8. doi:10.2460/javma.23.07.0372
- 20.↑
Bennett NE, Johnson EA, Gray PB. Veterinary care providers recognize clinical utility of genetic testing but report limited confidence in interpreting direct-to-consumer results. Am J Vet Res. 2024;86(9):ajvr.24.09.0265. doi:10.2460/ajvr.24.09.0265
- 21.↑
Leighton JW, Valverde K, Bernhardt BA. The general public's understanding and perception of direct-to-consumer genetic test results. Public Health Genomics. 2011;15(1):11–21. doi:10.1159/000327159
- 22.↑
Gunter LM, Barber RT, Wynne CDL. A canine identity crisis: genetic breed heritage testing of shelter dogs. PLoS One. 2018;13(8):e0202633. doi:10.1371/journal.pone.0202633
- 23.↑
Olson KR, Levy JK, Norby B, et al. Inconsistent identification of pit bull-type dogs by shelter staff. Vet J. 2015;206(2):197–202. doi:10.1016/j.tvjl.2015.07.019
- 24.↑
Hoffman CL, Harrison N, Wolff L, Westgarth C. Is that dog a pit bull? A cross-country comparison of perceptions of shelter workers regarding breed identification. J Appl Anim Welf Sci. 2014;17(4):322–339. doi:10.1080/10888705.2014.895904
- 25.↑
Voith VL, Trevejo R, Dowling-Guyer S, et al. Comparison of visual and DNA breed identification of dogs and inter-observer reliability. Am J Sociol Res. 2013;3(2):17–29.
- 26.↑
Voith VL, Ingram E, Mitsouras K, Irizarry K. Comparison of adoption agency breed identification and DNA breed identification of dogs. J Appl Anim Welf Sci. 2009;12(3):253–262. doi:10.1080/10888700902956151
- 27.↑
Visual breed identification. National Canine Research Council. Updated March 10, 2021. Accessed January 24, 2025. https://www.nationalcanineresearchcouncil.com/visual-breed-identification/
- 28.↑
Dreger DL, Hooser BN, Hughes AM, et al. True colors: commercially-acquired morphological genotypes reveal hidden allele variation among dog breeds, informing both trait ancestry and breed potential. PLoS One. 2019;14(10):e0223995. doi:10.1371/journal.pone.0223995
- 29.↑
Lindsay SR. Handbook of Applied Dog Behavior and Training. Vol 1. Iowa State University Press; 2000.
- 30.↑
Rajpurkar P, Chen E, Banerjee O, Topol EJ. AI in health and medicine. Nat Med. 2022;28(1):31–38. doi:10.1038/s41591-021-01614-0
- 31.↑
Johnson KB, Wei W, Weeraratne D, et al. Precision medicine, AI, and the future of personalized health care. Clin Transl Sci. 2020;14(1):86–93. doi:10.1111/cts.12884
- 32.↑
Dastin J. Amazon scraps secret AI recruiting tool that showed bias against women. Reuters. 2018. Accessed January 17, 2025. https://www.reuters.com/article/world/insight-amazon-scraps-secret-ai-recruiting-tool-that-showed-bias-against-women-idUSKCN1MK0AG
- 33.↑
Gigerenzer G, Gaissmaier W. Heuristic decision making. Annu Rev Psychol. 2011;62:451–482. doi:10.1146/annurev-psych-120709-145346
- 34.↑
De Neys W, Goel V. Heuristics and biases in the brain: dual neural pathways for decision making. In: Vartanian O, Mandel DR, eds. Neuroscience of Decision Making. Psychology Press; 2011:125–142.
- 35.↑
Herzog H. Forty-two thousand and one Dalmatians: fads, social contagion, and dog breed popularity. Soc Anim. 2006;14(4):383–397. doi:10.1163/156853006778882448
- 36.↑
Vakali A, Tantalaki N. Rolling in the deep of cognitive and AI biases. arXiv. July 30, 2024. Accessed October 19, 2024. https://arxiv.org/abs/2407.21202v1
- 37.↑
Rajhans A, Gupta R, Dubey S, Dave P, Patil A. Dog breed prediction using deep learning and E-commerce. IJSRCSEIT. 2023;9(2):446–451. doi:10.32628/CSEIT251445
- 38.↑
Sinnott RO, Wu F, Chen W. A mobile application for dog breed detection and recognition based on deep learning. In: Proceedings of the 2018 IEEE/ACM 5th International Conference on Big Data Computing Applications and Technologies (BDCAT); December 17–20, 2019; Zurich, Switzerland. doi:10.1109/BDCAT.2018.00019
- 39.↑
Everingham M, Van Gool L, Williams CKI, Winn JM, Zisserman A. The PASCAL visual object classes (VOC) challenge. Int J Comput Vis. 2009;88(2):303–338. doi:10.1007/s11263-009-0275-4
- 40.↑
Everingham M, Eslami SMA, Van Gool L, Williams CKI, Winn J, Zisserman A. The PASCAL visual object classes challenge: a retrospective. Int J Comput Vis. 2014;111:98–136. doi:10.1007/s11263-014-0733-5
- 41.↑
Wang Y, Wang Z. A survey of recent work on fine-grained image classification techniques. J Vis Commun Image Represent. 2019;59:210–214. doi:10.1016/j.jvcir.2018.12.049
- 42.↑
Zhao B, Feng J, Wu X, Yan S. A survey on deep learning-based fine-grained object classification and semantic segmentation. Int J Autom Comput. 2017;14(2):119–135. doi:10.1007/s11633-017-1053-3
- 43.↑
Khosla A, Jayadevaprakash N, Yao B, Li FF. Novel dataset for fine-grained image categorization: Stanford dogs. MIT Computer Science & Artificial Intelligence Laboratory. 2011. Accessed October 12, 2024. https://people.csail.mit.edu/khosla/papers/fgvc2011.pdf
- 44.↑
Valarmathi B, Gupta NS, Prakash G, Reddy RH, Saravanan S, Shanmugasundaram P. Hybrid deep learning algorithms for dog breed identification—a comparative analysis. IEEE Access. 2023;11(1):77228–77239. doi:10.1109/ACCESS.2023.3297440
- 45.
Borwarnginn P, Kusakunniran W, Karnjanapreechakorn S, Thongkanchorn K. Knowing your dog breed: identifying a dog breed with deep learning. Int J Autom Comput. 2020;18(1):45–54. doi:10.1007/s11633-020-1261-0
- 46.↑
Zou DN, Zhang SH, Mu TJ, Zhang M. A new dataset of dog breed images and a benchmark for finegrained classification. Comp Vis Media. 2020;6(1):477–487. doi:10.1007/s41095-020-0184-6
- 47.
Mondal A, Samanta S, Jha V. A convolutional neural network-based approach for automatic dog breed classification using modified-xception model. In: Lecture Notes in Electrical Engineering. Springer Nature Singapore; 2022;61–70.
- 48.↑
Raduly Z, Sulyok C, Vadaszi Z, et al. Dog breed identification using deep learning. In: Proceedings of the 2018 IEEE 16th International Symposium on Intelligent Systems and Informatics (SISY); September 13–15, 2018; Subotica, Serbia. doi:10.1109/SISY.2018.8524715
- 49.↑
Deng J, Dong W, Socher R, et al. ImageNet: a large-scale hierarchical image database. In: Proceedings of the 2009 IEEE Conference on Computer Vision and Pattern Recognition; June 20–25, 2009; Miami, FL.
- 51.↑
Dog breed identification. Kaggle. Accessed January 24, 2025. https://www.kaggle.com/c/dog-breed-identification/overview
- 52.↑
Liu J, Kanazawa A, Jacobs D, Belhumeur P. Dog breed classification using part localization. In: Fitzgibbon A, Lazebnik S, Perona P, Sato Y, Schmid C, eds. Lecture Notes in Computer Science. Springer Berlin Heidelberg; 2012:172–185.
- 53.↑
Mullen KR, Tammen I, Matentzoglu NA, et al. The vertebrate breed ontology: towards effective breed data standardization. Accessed October 17, 2024. https://arxiv.org/abs/2406.02623
- 54.↑
Parker HG, Dreger DL, Rimbault M, et al. Genomic analyses reveal the influence of geographic origin, migration, and hybridization on modern dog breed development. Cell Rep. 2017;19(4):697–708. doi:10.1016/j.celrep.2017.03.079
- 55.↑
Shannon LM, Boyko RH, Castelhano M, et al. Genetic structure in village dogs reveals a Central Asian domestication origin. Proc Natl Acad Sci U S A. 2015;112(44):13639–13644. doi:10.1073/pnas.1516215112
- 56.↑
Zhang H, Chen L. The complete mitochondrial genome of dhole Cuon alpinus: phylogenetic analysis and dating evolutionary divergence within Canidae. Mol Biol Rep. 2010;38(3):1651–1660. doi:10.1007/s11033-010-0276-y
- 57.↑
Koepfli KP, Pollinger J, Godinho R, et al. Genome-wide evidence reveals that African and Eurasian golden jackals are distinct species. Curr Biol. 2015;25(16):2158–2165. doi:10.1016/j.cub.2015.06.060
- 58.↑
Campana MG, Parker LD, Hawkins MTR, et al. Genome sequence, population history, and pelage genetics of the endangered African wild dog (Lycaon pictus). BMC Genomics. 2016;17(1):1013. doi:10.1186/s12864-016-3368-9
- 59.↑
Field MA, Yadav S, Dudchenko O, et al. The Australian dingo is an early offshoot of modern breed dogs. Sci Adv. 2022;8(16):eabm5944. doi:10.1126/sciadv.abm5944
- 60.↑
Cairns KM, Crowther MS, Parker HG, Ostrander EA, Letnic M. Genome-wide variant analyses reveal new patterns of admixture and population structure in Australian dingoes. Mol Ecol. 2023;32(15):4133–4150. doi:10.1111/mec.16998
- 61.↑
贵宾犬和泰迪的区别, 看了这个, 我终于明白了. 知乎专栏. Accessed January 24, 2025. https://zhuanlan.zhihu.com/p/292902141
- 62.↑
Li B. 泰迪犬,其实我就是贵宾犬 (Teddy dog, actually I am a Poodle). Zhongguo Gongzuo Quanye (China Working Dog) 2012;11:55–57. Available at: https://www.cnki.net/KCMS/detail/detail.aspx?dbcode=CJFD&dbname=CJFD2012&filename=GZQY201211019&uniplatform=OVERSEA&v=CGRVb-_OqjMd6jg03RuuuK4u499g5vfNGzuRvTvvCL3PN54EQtZG0c56DLgiIfQS
- 63.↑
Geiger RS, Cope D, Ip J, et al. “Garbage in, garbage out” revisited: what do machine learning application papers report about human-labeled training data? Quant Sci Stud. 2021;2(3):795–827. doi:10.1162/qss_a_00144
- 64.↑
Benthall S, Haynes BD. Racial categories in machine learning. In: Proceedings of the Conference on Fairness, Accountability, and Transparency. ACM, 2019;289–298. Available at: gftmpt. doi:10.1145/3287560.3287575
- 65.↑
Završnik A. Criminal justice, artificial intelligence systems, and human rights. ERA Forum. 2020;20:567–583. doi:10.1007/s12027-020-00602-0
- 66.↑
Königs P. Government surveillance, privacy, and legitimacy. Philos Technol. 2022;35(1):8. doi:10.1007/s13347-022-00503-9
- 67.↑
Fontes C, Hohma E, Corrigan CC, et al. AI-powered public surveillance systems: why we (might) need them and how we want them. Technol Soc. 2022;71:102137. doi:10.1016/j.techsoc.2022.102137
- 68.↑
Wang C, Wang J, Du Q, Yang X. Dog breed classification based on deep learning. In: Proceedings of the 2020 13th International Symposium on Computational Intelligence and Design (ISCID); December 12–13, 2020; Hangzhou, China. doi:10.1109/ISCID51228.2020.00053
