Artificial Intelligence in Chest Radiography—A Comparative Review of Human and Veterinary Medicine

Alessandro Rubini, Roberto Di Via, Vito Paolo Pastore, Francesca Del Signore, Martina Rosto, Andrea De Bonis, Francesca Odone, Massimo VignoliVeterinary 2025 · DOI

TL;DR

This review compares the application of Artificial Intelligence (AI) in chest radiography (CXR) for both human and veterinary medicine. While AI has shown significant promise in enhancing diagnostic speed and accuracy in human healthcare, its adoption in veterinary medicine is still nascent, facing challenges like limited datasets and breed variability. The paper highlights AI's potential as a supportive tool, emphasizing the continued need for human expertise in both fields.

What Problem Does This Paper Address?

The paper addresses the growing integration of Artificial Intelligence (AI) into medical imaging, specifically chest radiography (CXR), and aims to provide a comparative overview of its application and potential in both human and veterinary medicine. The core challenge lies in understanding how AI can effectively augment diagnostic capabilities, improve efficiency, and overcome limitations in image interpretation. In human medicine, AI is already being used to assist in identifying various thoracic abnormalities, improving emergency diagnostics, and classifying complex conditions. However, issues like data limitations, algorithmic biases, and ethical concerns still necessitate human oversight. In veterinary medicine, AI's application is even more limited due to smaller datasets, significant anatomical variations across breeds, and a general lack of extensive research. This paper seeks to bridge this gap by comparing the current state, challenges, and future directions of AI in CXR for both domains.

Methods

This review adopts a critical and interpretative approach, analyzing recent and highly cited applications of AI in CXR for both human and veterinary medicine. It does not strictly follow a systematic review model but aims to provide a reasoned overview of AI tools by selecting clinically relevant studies. The methodology involves comparing the similarities and differences between the two fields in terms of methodological approaches and performance. Particular attention is given to studies that not only present technical innovations but also offer practical insights into their applicability in real-world clinical settings. The review identifies key trends, potential discrepancies in outcomes, and sector-specific challenges to highlight differences and promote a collaborative understanding for future research and development.

Datasets

The paper discusses various datasets used in AI research for CXR interpretation.

  • Human Medicine: Large, well-structured datasets are frequently mentioned, such as the NIH CXR14 dataset, which is noted for its size but also for potential biases. Datasets for specific conditions like tuberculosis, pneumonia, and COVID-19 are also referenced.
  • Veterinary Medicine: Datasets are generally described as smaller and more limited compared to human medicine. Examples include datasets of canine chest radiographs for detecting left atrial enlargement (792 images), cardiogenic pulmonary edema (500 canine chest radiographs), and pleural effusion (62 dogs). Studies also mention the use of large datasets of canine and feline radiographs for general abnormality detection (approximately 2.5 million images).

The paper highlights that the size and quality of these datasets are critical factors influencing AI model performance and generalizability, particularly in veterinary medicine where obtaining large, representative datasets is a significant challenge.

Key Insights and Ideas

  • AI as a Powerful Augmentation Tool: Across both human and veterinary medicine, AI is presented not as a replacement for clinicians but as a valuable support tool that can enhance diagnostic speed, accuracy, and confidence.
  • Human Expertise Remains Crucial: Despite AI's advancements, human clinical judgment is indispensable for contextualizing AI findings, identifying subtle errors, and making final diagnostic decisions.
  • Data is King (Especially in Veterinary Medicine): The performance and reliability of AI models are heavily dependent on the quality, size, and representativeness of the training data. Veterinary medicine faces a significant hurdle due to the limited availability of large, standardized datasets compared to human medicine.
  • Breed Variability is a Major Hurdle in Veterinary AI: The anatomical differences and variability among different dog breeds pose a substantial challenge for developing generalized AI models in veterinary CXR interpretation. Focusing on species with less phenotypic variability, like cats, is suggested as a potential strategy.
  • Ethical and Practical Considerations: Beyond technical performance, ethical concerns (data privacy, algorithmic bias) and practical limitations (computational resources, regulatory hurdles) are significant factors influencing AI adoption in both fields.
  • Transfer Learning Potential: Research suggests that pre-training AI models on large natural image datasets (like ImageNet) might be as effective as in-domain pre-training for certain tasks like anatomical landmark detection in X-rays, which could be beneficial for veterinary applications where in-domain datasets are scarce.

Key Results

  • Human Medicine:
    • Deep learning algorithms have demonstrated significant improvements in identifying abnormalities like tuberculosis, malignant nodules, and COVID-19 pneumonia.
  • One DL algorithm (DLAD-10) correctly classified critical abnormalities with higher accuracy (95.0%) than radiologists (84.4%).
  • Radiologists assisted by DL algorithms showed significantly better reading performance, with higher AUC values (0.808 vs. 0.713).
  • AI has shown potential in early detection of heart failure, with AUCs around 0.798.
  • Veterinary Medicine:
    • CNN models achieved high accuracy (92.9% to 96.9%) and sensitivity (92.1% to 100%) in analyzing canine CXR images for abnormalities like cardiac silhouette and lung parenchyma.
  • Commercial AI software achieved accuracy of 92.3% for detecting cardiogenic pulmonary edema in dogs, with a high negative predictive value (99%).
  • Studies on pleural effusion detection in dogs showed accuracies around 88.7%.
  • AI models for detecting pulmonary nodules and masses in dogs showed lower accuracy (69.3%) but still demonstrated potential as a clinical support tool.

Limitations and Future Work

  • Limitations:
    • Data Limitations: The primary limitation in veterinary medicine is the scarcity of large, well-annotated datasets, which impacts model robustness and generalizability.
  • Breed and Species Variability: Significant anatomical differences between breeds and species in veterinary medicine present a challenge for AI model development.
  • Algorithmic Bias: AI models can inherit biases from training data, leading to potential misdiagnoses, especially in specific patient populations.
  • "Black Box" Problem: The lack of interpretability in deep learning models makes it difficult to understand the reasoning behind AI decisions, potentially eroding trust.
  • Limited Validation: Many veterinary AI studies use smaller datasets and may lack extensive validation, hindering widespread clinical adoption.
  • Human Oversight Still Necessary: AI is prone to errors that require human clinical judgment for identification and contextualization.
  • Future Work:
    • Focused Research on Less Variable Species: Investigating AI applications in species with less phenotypic variability, such as cats, could facilitate more accurate model training.
  • Interdisciplinary Collaboration: Encouraging collaboration between human and veterinary medicine researchers can foster knowledge transfer and accelerate AI development.
  • Data Augmentation and Sharing: Strategies for augmenting and sharing veterinary datasets are crucial to bridge the disparity with human medicine.
  • Developing Explainable AI (XAI): Research into more interpretable AI models is needed to build trust and facilitate clinical integration.
  • Standardization and Validation: More rigorous validation studies and standardization of AI applications are required for clinical implementation.
  • Exploring Transfer Learning: Leveraging pre-trained models from human medicine and adapting them for veterinary applications could be a promising avenue.
  • Focus on Clinical Relevance: Future research should prioritize clinically relevant applications that offer practical benefits to veterinary practitioners.

Summary updated

Abstract

The integration of artificial intelligence (AI) into chest radiography (CXR) has greatly impacted both human and veterinary medicine, enhancing diagnostic speed, accuracy, and efficiency. In human medicine, AI has been extensively studied, improving the identification of thoracic abnormalities, diagnostic precision in emergencies, and the classification of complex conditions such as tuberculosis, pneumonia, and COVID-19. Deep learning-based models assist radiologists by detecting patterns, generating probability maps, and predicting outcomes like heart failure. However, AI is still supplementary to clinical expertise due to challenges such as data limitations, algorithmic biases, and the need for extensive validation. Ethical concerns and regulatory constraints also hinder full implementation. In veterinary medicine, AI is still in its early stages and is rarely used; however, it has the potential to become a valuable tool for supporting radiologists in the future. However, challenges include smaller datasets, breed variability, and limited research. Addressing these through focused research on species with less phenotypic variability (like cats) and cross-sector collaborations could advance AI in veterinary medicine. Both fields demonstrate AI's potential to enhance diagnostics but emphasize the ongoing need for human expertise in clinical decision making. Differences in anatomy structure between the two fields must be considered for effective AI adaptation.
BibTeX
@inproceedings{rubini2025artificial,
  author = {Alessandro Rubini and Roberto Di Via and Vito Paolo Pastore and Francesca Del Signore and Martina Rosto and Andrea De Bonis and Francesca Odone and Massimo Vignoli},
  title = {{Artificial Intelligence in Chest Radiography—A Comparative Review of Human and Veterinary Medicine}},
  booktitle = {Veterinary},
  year = {2025},
  doi = {10.3390/vetsci12050404},
}
Figure