Explainable machine learning for assessing upper respiratory tract of racehorses from endoscopy videos.

The research proposes an efficient method to assess upper respiratory tract complications in racehorses using explainable machine learning to analyze endoscopy videos. The system uses a specific architecture for image segmentation, feature extraction, and to classify conditions severity, and has shown promising results on test cases.

Problem Background

• Laryngeal hemiplegia (LH) is a major concern in the racing horse industry, affecting the upper respiratory tract (URT) of the animals. This condition is usually diagnosed and graded in terms of severity using endoscopy videos of the horse’s throat.
• Current methods of manual assessment of these videos has several limitations, including the poor quality of the videos and potential subjectivity of the grading framework.

Proposed Solution

• The researchers propose an explainable machine learning (ML) solution. This approach uses a cascaded YOLOv8 (You Only Look Once) architecture to segment key semantic regions and landmarks within each frame of the video.
• The system then extracts spatiotemporal features from these key landmarks and uses a decision tree model to classify the severity of the LH as Grade 1,2,3, or 4 – representing no LH, mild LH, moderate LH, and severe LH, respectively.

Validation of the Proposed Method

• The machine learning method was validated using a 5-fold cross-validation on a collection of 107 videos. The results were promising, with high sensitivity values for the four severity grades.
• This validation was extended to an external dataset of 72 cases, further confirming the system’s ability to generalize across different cases, again with high sensitivity values.

Explainability Features

• The proposed system includes several features for explainability such as visualization of the YOLOv8 output to detect landmark estimation errors, time-series visualization to assess video quality and backtracking of the decision tree output to identify borderline cases.
• The incorporation of veterinarian diagnostic procedures into the machine learning framework ensures that the tool has clinical relevance and can complement the work of veterinarians by making the URT assessment process quicker and more reliable.