TICD: a novel thermal imaging cats' dataset for non-invasive health monitoring

1 Introduction

Digital health technologies have revolutionized the healthcare landscape by providing non-invasive, efficient, and scalable solutions for disease monitoring and early detection. Through the integration of sensors, imaging tools, and artificial intelligence (AI), digital health tools facilitated continuous and remote assessment of physiological and pathological conditions, resulting in better outcomes, lower costs, and improved patient compliance (1).

Among these tools, thermal imaging stands out for its ability to capture heat distribution across the body, offering insights into underlying physiological changes such as inflammation, infection, or circulatory irregularities. Unlike conventional diagnostic methods, thermal imaging is contactless, safe, and real-time, making it particularly useful in scenarios where minimizing stress is critical (2, 3).

In animal digital healthcare, thermal imaging has emerged as a promising tool for monitoring and diagnosing a range of conditions. Research has shown its utility in identifying musculoskeletal injuries in horses (4), mastitis in dairy cows (5, 6), and inflammation or fever in dogs (6). The authors in (7) reviewed the applications of thermal imaging in biomedical research for animals. The study presents that infrared thermography (IRT) can evaluate tissue viability, identify inflammation areas, and detect burn injuries in animals. In another review (8) investigated the use of thermal image analysis for disease detection in livestock. They suggest that IRT can detect diseases that affect the blood supply or cause inflammation, such as digital dermatitis, laminitis, and foot-and-mouth disease (FMD). They also highlight the inconsistencies in practices and the use of variable parameters in existing studies in the literature related to the use of thermography for animal disease detection. The non-invasive nature of thermal images is particularly beneficial for pets, as it minimizes stress and improves compliance during monitoring procedures.

Cats, in particular, present unique diagnostic challenges due to their subtle signs of illness and tendency to mask symptoms. The use of thermal imaging in cats' health monitoring holds strong potential, as it enables pet owners to detect localized heat changes indicative of injury, infection, or systemic illness without the need for visiting veterinary doctors.

1.1 Gap analysis

While thermal imaging has gained attention in veterinary diagnostics, there is a noticeable lack of publicly available datasets focusing on feline health assessment. Existing datasets and research primarily focus on livestock or dogs, leaving a gap in tools and data tailored to cats. To our knowledge, there is currently no open-access dataset specifically comprising thermal images of cats annotated for health monitoring purposes. This scarcity limits the development and validation of AI models aimed at the automatic detection or classification of feline health conditions using thermal data.

1.2 Contribution

To address this gap, we introduce TICD (Thermal Imaging Cats' Dataset), a novel dataset designed for non-invasive health monitoring of cats using thermal and digital imaging. TICD consists of 1,899 thermal and 527 digital images collected from 94 cats across various locations in Saudi Arabia, covering both indoor (controlled) and outdoor (uncontrolled) environments.

The dataset includes 81 images of healthy cats and 13 of symptomatic cats, providing a foundation for binary classification (healthy vs. sick) and potential future modeling of specific conditions. The digital images complement the thermal ones by documenting visible anomalies such as posture, wounds, or swelling.

The difference in the number of thermal images (1,899) vs. digital images (527) in the TICD dataset reflects their distinct purposes. Thermal images are the primary data source for analysis and model training. It is essential to collect more than one thermal image, as signs of heat may appear partially from one angle and more clearly from another. In contrast, digital images are included for documentation. Each digital image visually confirms the physical condition of the corresponding cat and provides a reference that links the thermal signature to the cat's visible appearance.

TICD provides researchers with a valuable resource for developing AI-driven diagnostic tools, benchmarking classification models, and conducting comparative analyses across various environmental conditions. By making TICD available, we aim to foster advancements in feline digital health and encourage further investigation into non-invasive diagnostic technologies.

1.3 Report organization

The remainder of this report is organized as follows: Section 2 presents the Materials and Methods used in creating the TICD dataset, including the methodology, data collection protocol, instrumentation, and dataset annotation. Section 3 demonstrates the potential of TICD for developing health classification models for cats. Section 4 provides a TICD baseline implementation and results for initial validation. Section 5 presents a discussion of the dataset's key aspects.

2 Materials and methods

2.1 Methodology

The process of constructing the TICD dataset followed a structured pipeline as shown in Figure 1A. The methodology involved selecting diverse collection locations, following ethical imaging procedures, and capturing images from multiple angles using thermal and digital cameras. These images were gathered across both controlled and uncontrolled settings to ensure environmental diversity and robustness in dataset composition.

Figure 1

Figure 1. The dataset organization and its utilization (A) represents the structure of the dataset. (B) Shows an example of the utilization of the dataset in classification.

2.2 Data collection protocol

We collected data over a six-month period, from September 2024 to March 2025, across two environmental settings:

• Controlled indoor environments such as veterinary clinics (Nabd Al Farah, Al-Asr Veterinary Doctors Center) and pet-related facilities (Black Forest Pets and Supplies, Cat Lounge).

• Uncontrolled outdoor environments in public areas of Al Ruwais (Jeddah) and Al-Usbah District (Medina), which varied in lighting, ambient temperatures, and background thermal noise.

Each cat image was taken from multiple angles using both traditional digital cameras and thermal imaging.

2.3 Instrumentation and setup

The thermal images in the TICD dataset were captured using a Seek Shot thermal camera with a resolution of 640 × 480 pixels, capable of capturing both thermal and digital images. The camera has a temperature sensitivity range of −40°C to 330°C, making it suitable for non-invasive surface temperature measurements. Thermal images are saved in JPEG format, embedded with thermal color maps representing temperature gradients in Celsius—red/yellow hues indicating warmer regions and blue hues indicating cooler areas.

To ensure the safety and comfort of the animals, we captured images from 90 cm to one meter. Given the natural behavior of cats—who tend to move frequently and unpredictably—images were taken from multiple angles. This variability was intentional, aiming to simulate real-world, non-laboratory conditions and enhance the robustness of machine learning models trained on the dataset. While we maintained as much consistency as possible in the imaging setup, some minor variations in lighting and angles were present. These variations reflect the dataset's goal of supporting practical and flexible thermal imaging applications. In parallel with the thermal imaging, high-resolution RGB digital images were captured to document visible physical conditions such as abnormal posture, wounds, or fur loss. These digital images serve as complementary data, providing visual cues that aid in the annotation, diagnosis, and validation of the thermal data.

2.4 Dataset annotation

Each cat in the dataset is assigned a unique identifier linking its digital and thermal images. The following metadata were recorded for each entry:

• Image type (thermal or digital),

• Health state (healthy or sick),

• Environmental setting (controlled or uncontrolled),

• Time of capture (day or night for uncontrolled settings).

The health classification (healthy vs. sick) was confirmed by licensed veterinarians based on clinical assessments. The final dataset includes 2,426 images:

• 1,899 thermal images

• 527 digital images

A full description of the TICD is explain in Table 1A, while Table 1B shows the full details of the dataset and how every cat is assigned an ID and shows all its relevant statistics.

Table 1

Table 1. The upper part (A) shows the specification and full description of the dataset, while the lower part (B) shows full details of the dataset and how every cat is assigned an ID and shows all its relevant statistics.

3 Potential of the TICD

The dataset supports downstream applications such as segmentation, temperature-based feature extraction, and classification of health states (healthy vs. sick) using machine learning and deep learning models (e.g., CNNs).

Figure 1B illustrates a workflow for classifying the health status of a cat (sick or healthy) using the TICD dataset. The process leverages image processing techniques combined with machine learning (ML) or deep learning (DL) classifiers. The workflow outlines the training process for a cat health classification model using thermal images from the TICD dataset. The workflow begins with a thermal image of a cat, which undergoes segmentation to isolate the animal from the background. A temperature histogram is generated from segmented image, providing key thermal statistics (average, min, max, most frequent temperature). These statistics are then compiled in a feature vector.

This features vector (or potentially the segmented image directly, if a CNN is used) is fed into an ML/DL classifier (e.g., a Deep Learning Classifier like a CNN). During the training phase, this classifier learns to associate specific temperature patterns or image features with the corresponding health state (sick or healthy) of the cat. The goal of this training is to build a robust model that can accurately predict a cat's health status from new thermal images.

4 TICD baseline

To establish a baseline for future research and provide a lower bound for comparison, we implemented the pipeline illustrated in Figure 1B. In this approach, the input thermal image undergoes a series of processing steps: segmentation, temperature histogram generation, feature vector extraction, and finally, classification using a machine learning model, logistic regression in our case. The model then outputs the predicted class labels.

The segmentation technique utilizes region masking and binary thresholding. First, specific areas such as the color bar, borders, and irrelevant regions are masked out using Boolean arrays. The remaining region is then converted to grayscale and thresholded to create a binary mask that isolates the cat's body. This segmented region is subsequently used to extract relevant thermal data, minimizing interference from background noise or overlay graphics.

After segmentation, the histogram generation module uses OCR to extract temperature scale values from the color bar, maps pixel colors to temperature values using linear interpolation, and calculates key statistical features. These include average temperature, minimum temperature, maximum temperature, and most frequent temperature, which serve as informative indicators of potential thermal anomalies related to cat's health status.

The dataset is divided into training and testing sets, with 80% allocated for training and 20% for testing. The logistic regression model achieved an accuracy of 73.42%, a specificity of 75.96%, a precision of 67.88%, a recall of 69.92%, and an F1 score of 68.89%.

5 Discussion

The TICD dataset includes only two labels: Sick and Healthy. The Sick label serves as a general indicator of illness, without specifying the cause, specific health condition, or symptoms. In its current form, TICD only indicates whether a cat is sick or not. The health labels in the TICD dataset are assigned based on joint confirmation by both the cat's owner and a licensed veterinarian. In all cases labeled as sick, the owner first reported signs of illness or abnormal behavior, which is clinically evaluated and confirmed by the veterinarian. This binary classification of Sick or Healthy represents a basic or initial level of investigation. In future iterations of the dataset, it can be extended to include more detailed information, such as the specific disease diagnosis and accompanying symptoms.

Although the TICD dataset includes 81 healthy cats and only 13 sick cats, resulting in a total of 1,899 thermal images, sick cats are captured more frequently per individual, contributing to 44.44% of the images, while healthy cats account for 55.56%. Despite this more balanced image-level distribution, the underlying class imbalance at the individual (cat) level can still impact machine learning performance. Specifically, models trained on such imbalanced data may become biased toward the majority class (healthy), leading to reduced sensitivity and a higher rate of false negatives when identifying sick cases. This issue especially concerns health-related applications, where accurately detecting abnormal or pathological conditions is critical. To help researchers address this imbalance, we recommend the following strategies:

• Class Rebalancing: Applying techniques such as under-sampling the majority class or over-sampling the minority class can help balance the dataset and improve model sensitivity to underrepresented classes.

• Synthetic Data Augmentation: Generating new samples for the minority class using methods like SMOTE (Synthetic Minority Over-sampling Technique) or employing image augmentation techniques (e.g., rotation, translation, flipping) can enhance data diversity and help models generalize better.

• Do not rely on accuracy alone. Use metrics that reflect performance on the minority class, like Precision, Recall, F1-score.

• Ensemble Methods: Using ensemble models that are more robust to imbalance classes like: Balanced Random Forest, XGBoost and LightGBM allow custom class weights.

• By applying these strategies, researchers can develop more balanced and reliable classification models, making full use of the TICD dataset's potential for non-invasive health monitoring of cats.

Data availability statement

The datasets presented in this study can be found in online repositories. The names of the repository/repositories and accession number(s) can be found below: https://doi.org/10.6084/m9.figshare.28926584.

Ethics statement

The animal studies were approved by Research Commitee of Faculty of Computer and Information System, Islamic University of Madinah, KSA. The studies were conducted in accordance with the local legislation and institutional requirements. Written informed consent was obtained from the owners for the participation of their animals in this study.

Author contributions

MA: Conceptualization, Investigation, Writing – review & editing, Writing – original draft, Data curation, Project administration, Resources, Validation, Methodology. AN: Writing – review & editing, Methodology, Supervision, Writing – original draft, Conceptualization, Project administration, Funding acquisition. AA: Validation, Writing – review & editing, Methodology, Data curation, Formal analysis, Writing – original draft, Investigation, Software. HG: Investigation, Methodology, Validation, Writing – review & editing, Software, Conceptualization, Writing – original draft. SA: Data curation, Writing – original draft, Validation, Resources, Writing – review & editing. EN: Investigation, Supervision, Methodology, Writing – review & editing, Writing – original draft.

Funding

The author(s) declare that financial support was received for the research and/or publication of this article. This research is funded by the Deanship of Scientific Research, Islamic University of Madinah, Madinah, Saudi Arabia.

Conflict of interest

The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

Generative AI statement

The author(s) declare that no Generative AI was used in the creation of this manuscript.

Any alternative text (alt text) provided alongside figures in this article has been generated by Frontiers with the support of artificial intelligence and reasonable efforts have been made to ensure accuracy, including review by the authors wherever possible. If you identify any issues, please contact us.

Publisher's note

All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.

Supplementary material

The Supplementary Material for this article can be found online at: https://www.frontiersin.org/articles/10.3389/fdgth.2025.1650223/full#supplementary-material

References

1. Steinhubl SR, Muse ED, Topol EJ. Can mobile health technologies transform health care? JAMA. (2013) 310(22):2395. doi: 10.1001/jama.2013.281078

PubMed Abstract | Crossref Full Text | Google Scholar

2. Lahiri BB, Bagavathiappan S, Jayakumar T, Philip J. Medical applications of infrared thermography: a review. Infrared Phys Technol. (2012) 55(4):221–35. doi: 10.1016/j.infrared.2012.03.007

PubMed Abstract | Crossref Full Text | Google Scholar

3. Nadeem A, Ashraf M, Mehmood A, Rizwan K, Siddiqui MS. Dataset of date palm tree (Phoenix dactylifera L.) thermal images and their classification based on red palm weevil (Rhynchophorus ferrugineus) infestation. Front Agron. (2025) 7:1604188. doi: 10.3389/fagro.2025.1604188

Crossref Full Text | Google Scholar

4. von Schweinitz DG. Thermographic diagnostics in equine back pain. Vet Clin N Am: Equine Pract. (1999) 15(1):161–77. doi: 10.1016/s0749-0739(17)30170-0

Crossref Full Text | Google Scholar

5. Sathiyabarathi M, Jeyakumar S, Manimaran A, Jayaprakash G, Pushpadass HA, Sivaram M, et al. Infrared thermography: a potential noninvasive tool to monitor under health status in dairy cows. Vet World. (2016) 9(10):1075–81. doi: 10.14202/vetworld.2016.1075-1081

PubMed Abstract | Crossref Full Text | Google Scholar

6. Mota-Rojas D, Martínez-Burnes J, Casas-Alvarado A, Gómez-Prado J, Hernández-Ávalos I, Domínguez-Oliva A, et al. Clinical usefulness of infrared thermography to detect sick animals: frequent and current cases. Cab Rev Perspect Agric Vet Sci Nutr Nat Resour. (2022) 17:1–17. doi: 10.1079/cabireviews202217040

Crossref Full Text | Google Scholar

7. Verduzco-Mendoza A, Olmos-Hernández A, Bueno-Nava A, Villanueva-García D, Domínguez-Oliva A, Avila-Luna A, et al. Thermal imaging in biomedical research: a non-invasive technology for animal models. Front Vet Sci. (2025) 12:1544112. doi: 10.3389/fvets.2025.1544112

PubMed Abstract | Crossref Full Text | Google Scholar

8. McManus R, Boden LA, Weir W, Viora L, Barker R, Kim Y, et al. Thermography for disease detection in livestock: a scoping review. Front Vet Sci. (2022) 9:965622. doi: 10.3389/fvets.2022.965622

PubMed Abstract | Crossref Full Text | Google Scholar