Classification of Pet Animals Skin Disease Using Vision Transformers

P. Gowsikraja, S. S. Samwilson, S. Joedinesh and K. Saran

Department of Computer Science and Design, Kongu Engineering College (Autonomous) Perundurai,

Erode - 638 060, Tamil Nadu, India

Keywords: Vision Transformers, Pet Skin Diseases, Deep Learning, Image Classification, Veterinary Diagnostics.

Abstract: The classification of pet skin diseases is becoming more and more necessary due to the increasing demand

for early and accurate diagnostics in veterinary healthcare. Complex and diverse skin disease patterns of pets

urge advanced models for reliable recognition and diagnosis. In this paper, we propose an approach based on

Vision Transformers (ViTs) for the classification of skin diseases in pets. The proposed model would be

capable of detecting and categorizing a number of skin diseases efficiently after processing images in such a

way that detailed features crucial for accurate classification are captured. The model has been trained on a

dataset of annotated images to better generalize and perform. These results will be of very important

significance in the diagnosis of different skin diseases found in pets and will contribute a lot to veterinary

dermatology while propagating modern deep learning techniques in improving diagnostic accuracy.

1 INTRODUCTION

It reflects a growing demand for veterinary care,

especially dermatology, due to the skin diseases in

pets that are very common. Therefore, these

conditions should be diagnosed early and correctly

classified so that proper treatment can be initiated,

hence overall health of the animals. However,

traditional methods used for diagnostics present an

unsuitable case as disease manifestation variability is

evident, even in front of seasoned veterinarians.

Techniques based on deep learning are indeed very

promising as they may well automate such diagnoses

and bring improved accuracy into it. A new

architecture in the computer vision domain, known as

Vision Transformers or ViTs, was surprisingly

effective for most classification tasks. This is due to

its ability to model long-range dependencies in

images. The paper is based on the application of ViTs

for the classification of pet skin disease. It elaborates

on how it applies, performs, and compares with other

traditional models, such as CNNs. In this study, the

aim is to design a robust model for distinguishing

various skin diseases in pets and to improve diagnostic

efficiency. Traditional techniques of diagnosis are

mostly based on observation, and hence, vary because

of the asymmetrical presentation of symptoms. This

brings out the emerging importance of AI-based

techniques to aid in support for diagnostic activities,

especially for image classification. Deep learning

architectures, especially of advanced form known as

Vision Transformers, have shown a lot of promise in

image analysis and classification that even includes

medical images. We will apply the application of ViTs

on a pet skin disease classification task and extend

veterinary healthcare diagnostics from a set of images

specially labeled for different kinds of skin diseases.

2 RELATED WORKS

Artificial intelligence, particularly deep learning, has

made significant strides in medical image

classification. Convolutional Neural Networks

(CNNs) have long been the dominant architecture in

image classification tasks, with numerous studies

demonstrating their effectiveness in identifying

human skin diseases such as melanoma and psoriasis.

These models excel at extracting hierarchical features

but can struggle when handling the complex patterns

often present in pet skin diseases.

The Vision Transformer architecture offers an

alternative to CNNs by treating images as a sequence

of patches, which allows the model to retain

information from across the entire image. Unlike

CNNs, which focus on local feature extraction through

convolutional filters, ViTs can process global image

information more effectively. This characteristic

888

Gowsikraja, P., Samwilson, S. S., Joedinesh, S. and Saran, K.

Classiﬁcation of Pet Animals Skin Disease Using Vision Transformers.

DOI: 10.5220/0013875300004919

Paper published under CC license (CC BY-NC-ND 4.0)

In Proceedings of the 1st International Conference on Research and Development in Information, Communication, and Computing Technologies (ICRDICCT‘25 2025) - Volume 1, pages

888-893

ISBN: 978-989-758-777-1

makes them well-suited for tasks where the

relationships between distant regions of an image are

important, such as in the classification of skin diseases

with varied presentations. While ViTs have shown

promise in fields like object detection and medical

imaging, their application in veterinary diagnostics

remains relatively unexplored.

3 METHODOLOGY

3.1 Dataset Preparation

Table 1: Dataset Distribution by Disease Type.

Disease Type

Number of

Images

Percentage

(%)

Healthy Skin 200 12%

Fungal Infections 350 21%

Bacterial Dermatosis 300 18%

Hypersensitivity

Allergic Dermatosis

350 21%

Other Conditions 100 6%

Total 1300 100

For this table 1, the dataset consists of images of three

primary skin disease categories in dogs namely Fungal

Infections, Hypersensitivity Allergic Dermatosis, and

Bacterial Dermatosis and includes data of healthy skin

to differentiate affected and healthier skins. These

images were collected from veterinary clinics and

medical repositories, ensuring that the dataset covers

a variety of symptoms and conditions for each disease

type.

• Fungal Infections: These include conditions

caused by fungi, such as ringworm,

characterized by scaly patches and hair loss.

• Hypersensitivity Allergic Dermatosis: This

refers to allergic reactions leading to skin

inflammation, itching, and swelling.

• Bacterial Dermatosis: These infections cause

redness, lesions, and, in some cases, pus

formation on the skin.

Each image in the dataset was manually labelled

by veterinary professionals to ensure accurate and

high-quality annotations. To prevent overfitting and

improve model generalization, data augmentation

techniques such as random rotations, scaling, and flips

were applied. This ensured that the model would

perform well across diverse scenarios.

The dataset was divided into three subsets: training

(70%), validation (15%), and testing (15%), with the

image resolution standardized to 224x224 pixels to fit

the input size requirements of the Vision Transformer

model.

3.2 Vision Transformer Model

This architecture based on ViT changes the processing

of an image in comparison with the standard CNNs.

The model does not rely on the usual convolutional

layers of the CNN to extract features but instead

functions by splitting the input image into non-

overlapping small patches. These are usually 16 × 16

pixels in size; then flattened to a one-dimensional

vector. It enables the model to take images as a

sequence, just like the way NLP tasks are addressed in

processing text. This embedding space allows these

vectors to be projected once the image is divided into

patches. In this context, the embedding becomes

crucial for allowing the model to understand

relationships among different patches. The

transformer layers are the backbone of the Vision

Transformer. These employ self-attention

mechanisms, which gives more relevance to some

patches with respect to the other patches, allowing the

model to capture more patterns and the correlations

between the local and the global features that exist in

the images.

The multi-head attention mechanism in the

transformer layers gives a deep, nuanced

understanding of content through focusing differently

at various places of the images at once. This is

especially the case for disease classification problems,

where sometimes slight differences in texture and

structure can signify different skin diseases. The

output of the transformer layers proceeds to a

classification head, often a Multi-Layer Perceptron

(MLP), to produce the final predictions for the

categories for the skin disease. See the architecture of

the vision transformer in Figure 1.

Classiﬁcation of Pet Animals Skin Disease Using Vision Transformers

889

Figure 1: Vision Transformer Architecture.

3.3 Model Specifications

The Vision Transformer model for this study was

configured with the following specifications:

• Input Image Size: 224×224 pixels. This size

is commonly used to maintain a balance

between computational efficiency and

sufficient detail for accurate classification.

• Patch Size: 16×16 pixels. This patch size

allows for capturing localized features while

maintaining a manageable number of patches

for processing.

• Number of Layers: 12 transformer layers.

The depth of the model helps it learn

complex representations.

• Embedding Dimension: 768. This high-

dimensional space allows for a more

expressive representation of the input data.

• Attention Heads: 12. Multiple attention

heads enable the model to attend to different

parts of the image, capturing various features

effectively.

• Classification Head: Multi-Layer

Perceptron (MLP), which processes the

aggregated information from the transformer

layers and outputs probabilities for each class

of skin disease.

3.4 Flow Chart

Figure 2: Flow Chart.

ICRDICCT‘25 2025 - INTERNATIONAL CONFERENCE ON RESEARCH AND DEVELOPMENT IN INFORMATION,

COMMUNICATION, AND COMPUTING TECHNOLOGIES

890

3.5 Training Process

The training process followed a stepwise procedure to

the Vision Transformer, thus ensuring correct learning

and proper generalization for unseen data. Using the

Adam optimizer was pertinent, considering that it

efficiently takes care of sparse gradients and adjusts

learning rates in real time. The chosen learning rate

was 0.0001, balanced to provide maximum

convergence speed within stable training processes.

Training has been performed in batches of 32 images;

this is ensured to provide sufficient computation with

regard to memory use. Batch size influences the

generalization capability of the model; too large a

batch size tends to generalize poorly while the

opposite may lead to noisy updates. The loss function

is cross-entropy that is well suited for multi-class

classification problems. This function calculates the

difference of class probabilities, which the model has

predicted with the actual class labels, and these

differences guide the model to reduce its prediction

errors. It is trained for 50 epochs with the early

stopping mechanisms, which mean that the training is

stopped if the model fails to improve for a specified

number of epochs on the validation set. This can

prevent overfitting in the models, a typical problem

that often arises in deep learning models when the

dataset is relatively small.

The above random rotations, scaling, flipping, and

colour augmentation are then used to build the model

as robust. The model learns it to be exposed to a lot

more scenarios than it would otherwise have been

under if all of these were to have been learned on live

data. Finally, there was transfer learning. Here the

initialization is with the weights attained by

pretraining on the ImageNet dataset. This approach

now puts the knowledge learned from a large-scale

dataset to good use: it now constitutes an extremely

strong initialization for the model. It would allow the

ViT model to adapt even better to the particular task

of classification of skin disease, even by being

compatible with a relatively small target dataset.

4 RESULT AND DISCUSSIONS

4.1 Performance Evaluation

Test set performance was conducted on the ViT model

by focusing considerably on major three key

performance metrics of evaluation-precision, recall,

F1-score, and accuracy. The overall accuracy rate

returned was 90.5%, showing how the model can

classify images very well according to different

categories of skin diseases. Besides this, all precision,

recall, and F1-score of the model were more than 90%

with respect to diagnostic accuracy, as well as

consistency in finding accuracy with the right disease.

The precision value refers to the accuracy with which

the model will classify its positive instances without

producing excessive false positives. It is very reliable

for diagnostics. The recall score shows how good the

model is in classifying true positives and, thus,

avoiding a possible misclassification. The high F1-

score combines precision and recall, indicating that

the model has the strength to handle the complexity

associated with variable presentations of the disease.

These results confirm reliability for models in real-

veterinary and actual medical application fields where

diagnoses are important as they have to be both precise

yet speedy. The ViT Vision transformer can view full

images and identify between two very similar

conditions so that such specifics appear well distinct.

In this success, therefore, deep learning models like

ViT are most likely to change practice in the sense that

they are going to make such a diagnosis. This is

concerning areas such as dermatology. Table 2

represents the model performance metrics and figure

3 and 4 shows the model evaluation metrices and

confusion matrix.

Table 2: Model Performance Metrics.

Metric Value

Accuracy 90.5%

Precision 91.2%

Recall 90.8%

F1- Score 91.0%

Figure 3: Model Evaluation Metrices.

Classiﬁcation of Pet Animals Skin Disease Using Vision Transformers

891

Figure 4: Confusion Matrix.

4.2 Discussion

The model proved to outperform on the different

classes of disease. It can extract a very wide-ranging

pattern in the entire image and proved to be way better

than approaches that highly depend on these local

features. Actually, it showed a really impressive

performance between different skin conditions which

look pretty similar, such as fungal infections versus

bacterial infections. Still, the scope to enhance the

performance further is still in place. Generalizing

capabilities of this model will also be remarkably high

if diversity, such as greater instances of uncommon

skin diseases or even broader species than cats and

dogs, of the dataset will be enhanced. This would

ensure that more precise diagnoses occur in real

veterinary practice as its capability to robustly deal

with different scenarios it might face would be

increased.

5 CONCLUSIONS

This research demonstrates the promising capabilities

of the ViTs in accomplishing the task of skin disease

classification in pets. The results show that it was

found to be accurate under different conditions, hence

really applicable to help veterinary diagnostics get

better. Application of ViTs would therefore greatly

enhance the accuracy as well as the speed of diagnosis,

which, in a way, assists in better planning of treatment

for pets, indirectly enhancing their quality of life. In

particular, the ability to pull complex patterns and

global features from images will be helpful in

differentiating presentations of skin disease that may

look very much alike. This feature of the ViTs will

prove to be of utmost value in veterinary practice,

thanks to the ability to make a diagnosis rapidly and

precisely, therefore arresting the course of disease and

allowing intervention sooner rather than later.

Add new data modalities, for example, clinical notes

or owner-reported symptoms and treatment history, in

order to have a holistic diagnostics approach that is

going to stretch the model very far. A multimodal

approach like this one would add visual context,

beside the mere analysis of vision; it increases the

accuracy of the classification more than just purely

visual analysis of the model above. It will also require

a dataset large and varied enough to cover different

skin conditions and species of pets so that the model

would be robust and generalizable in real-world

veterinary settings.

In summary, the successful application of Vision

Transformers in the present study suggests that these

have the potential to revolutionize the face of

veterinary diagnostics. With development in the

future, technologies of AI are bound to change how

veterinarians will handle diagnostics and treatment on

balance, better outcomes for the pets and less stress

among the owners.

REFERENCES

Bhavsar, S., & Mehendale, N. (2022). Deep Learning-

Based Automatic System for Diagnosis and

Classification of

Skin Dermatoses. https://doi.org/10.21203/rs.3.rs- 236

0579/v1

Gupta, P., & Gupta, S. (2022). Deep learning in medical

image classification and object detection: A survey.

International Journal of Image Processing and Pattern

Recognition. https://doi.org/10.37628/ijippr.v8i2.846

Himel, G. M., Islam, Md. M., Al-Aff, Kh. A., Karim, S. I.,

& Sikder, Md. K. (2024). Skin cancer segmentation and

classification using vision transformer for automatic

analysis in dermatoscopy-based noninvasive digital

system. International Journal of Biomedical Imaging,

2024, 1–18. https://doi.org/10.1155/2024/3022192

Hwang, S., Shin, H. K., Park, J. M., Kwon, B., & Kang, M.-

G. (2022). Classification of dog skin diseases using

deep learning with images captured from Multispectral

Imaging device. Molecular & Cellular

Toxicology, 18(3), 299–309.

https://doi.org/10.1007/s13273-022-00249-7

Hyeon Ki Jeong 1 2, 1, 2, & Artificial intelligence (AI) has

recently made great advances in image classification

and malignancy prediction in the field of dermatology.

However. (2022, August 23). Deep learning in

dermatology: A systematic review of current approach

-es, outcomes, and limitations. JID Innovations.

https://www.sciencedirect.com/science/article/pii/S26

67026722000583

Jiang, Z., Gu, X., Chen, D., Zhang, M., & Xu, C. (2024).

Deep learning-assisted multispectral imaging for early

ICRDICCT‘25 2025 - INTERNATIONAL CONFERENCE ON RESEARCH AND DEVELOPMENT IN INFORMATION,

COMMUNICATION, AND COMPUTING TECHNOLOGIES

892

screening of skin diseases. Photodiagnosis and

Photodynamic Therapy, 48, 104292. https://doi.org/10

.1016/j.pdpdt.2024.104292

Mohsin Ali, M., Chandra Joshi, R., & Kishore Dutta, M.

(2022). An automated and efficient deep learning-based

classification of multiple skin disorders from skin

lesion images. 2022 International Conference on Edge

Computing and Applications (ICECAA), 1156–

1161.https://doi.org/10.1109/icecaa55415.2022.99360

Rathnayaka, R. M. N. A., Anuththara, K. G. S. N.,

Wickramasinghe, R. J. P., Gimhana, P. S.,

Weerasinghe, L., & Wimalaratne, G. (2022). Intelligent

system for skin disease detection of dogs with ontology

based clinical information extraction. 2022 IEEE 13th

Annual Ubiquitous Computing, Electronics &

Mobile Communication Conference (UEMCON),

0059–0066.

https://doi.org/10.1109/uemcon54665.2022.9965696

S. P. R. R. Raj and P. K. Gupta, "Vision Transformers in

Medical Image Analysis: A Review," IEEE

Transactions on Medical Imaging, vol. 41, no. 9, pp.

2161- 2175, Sep. 2022. https://www.semanticscholar.

org/paper/Transformers-in-Medical-Image-Analysis:-

A-Review-He-

Gan/42bad1b72259aa1ff70d7ce2539220a83f1af9a4

Saraf, P., Tharaniesh, P. R., & Singh, S. (2024). Skin

disease detection using convolutional neural network.

2024 Second International Conference on Networks,

Multimedia and Information Technology (NMITCON),

1– 6. https://doi.org/10.1109/nmitcon62075.2024.106

99196

Upadhyay, A., Singh, G., Mhatre, S., & Nadar, P. (2023).

Dog skin diseases detection and identification using

convolutional neural networks. SN Computer Science,

4(3). https://doi.org/10.1007/s42979-022-01645-5

Classiﬁcation of Pet Animals Skin Disease Using Vision Transformers

893