# A Novel Machine Learning Model for the Automated Diagnosis of Nasal Pathology in Canine Patients

## TL;DR

This paper introduces a machine learning model that accurately classifies nasal diseases in dogs using CT scans, improving diagnostic accuracy in veterinary medicine.

## Contribution

A neural network-based pipeline for automated diagnosis of canine nasal pathologies using CT imaging is developed and validated.

## Key findings

- The model achieved 86% classification accuracy on isolated CT image slices.
- Aggregating results across slices improved diagnosis accuracy to 99%.
- The model can differentiate between normal anatomy, fungal rhinitis, and intranasal neoplasia.

## Abstract

Computed tomography (CT) is the modality of choice for assessing the canine nasal cavity, offering critical insights into disease extent, facilitating targeted tissue sampling, and informing therapeutic strategies. Although CT findings can provide indications of pathology type, considerable overlap exists among neoplastic, inflammatory, and infectious nasal diseases, complicating definitive differentiation. In human medicine, recent advancements in computer-aided detection have leveraged machine learning and deep learning techniques to enhance the identification and classification of intranasal pathology with high accuracy. This study aimed to develop a neural network-based pipeline for the automated detection and classification of nasal pathology in canines using CT imaging. A dataset comprising 80 CT studies of the head was curated for model training and validation. Each study was assigned to one of three categories—normal nasal anatomy, fungal rhinitis, or intranasal neoplasia—and manually segmented to train a series of neural networks. Performance was evaluated using standard accuracy metrics. The trained model demonstrated a classification accuracy of 86% on isolated image slices and a diagnosis accuracy of 99% when aggregated across slices of a given patient. These findings underscore the potential of machine learning algorithms in accurately differentiating intranasal pathologies in canines, highlighting their applicability in augmenting diagnostic workflows and advancing veterinary imaging.

Computed tomography (CT) is the imaging method of choice for evaluating the canine nasal cavity, being invaluable in determining disease extent, guiding sampling, and planning treatment. While predictions of pathology type can be made, there is significant overlap between CT changes noted in neoplastic, inflammatory, and infectious nasal disease. Recent years have seen remarkable advancement in computer-aided detection systems in human medicine, with machine and deep learning techniques being successfully applied for the identification and accurate classification of intranasal pathology. This study aimed to develop a neural network pipeline for differentiating nasal pathology in dogs using CT studies of the head. A total of 80 CT studies were recruited for training and testing purposes. Studies falling into one of the three groups (normal nasal anatomy, fungal rhinitis, and intranasal neoplasia) were manually segmented and used to train a suite of neural networks. Standard accuracy metrics assessed performance during training and testing. The machine learning algorithm showed reasonable accuracy (86%) in classifying the diagnosis from an isolated scan slice but high accuracy (99%) when aggregating over slices taken from a full scan. These results suggest that machine learning programmes can accurately discriminate between intranasal pathologies based on canine computed tomography.