Automated segmentation of canine pulmonary masses in CT imaging using AI

Area of Science:

• Veterinary Radiology
• Artificial Intelligence in Medicine
• Canine Oncology

Background:

• Primary lung cancer in dogs is uncommon, necessitating advanced imaging techniques for diagnosis.
• Manual segmentation of pulmonary lesions on CT scans is labor-intensive and prone to inter-observer variability.
• Accurate segmentation is crucial for treatment planning and monitoring canine lung cancer.

Purpose of the Study:

• To develop and evaluate an artificial intelligence (AI) model for automated segmentation of primary pulmonary masses in dogs.
• To assess the performance of the AI model using quantitative metrics on a multicenter dataset.
• To identify factors influencing the AI model's segmentation accuracy in canine lung cancer.

Main Methods:

• A retrospective multicenter dataset of 217 canine CT scans with pulmonary masses (>2 cm) was compiled.
• Lesions were manually segmented to create ground truth data.
• An AI model (nnUNet v2 framework) was trained and validated using 5-fold cross-validation, then tested on 30 scans.

Main Results:

• The AI model achieved high segmentation accuracy on the test set, with a mean Dice Similarity Coefficient (DSC) of 0.91.
• The mean Average Symmetric Surface Distance (ASSD) was 1.88 mm, indicating precise boundary delineation.
• Performance was optimal for homogeneous, well-defined masses; intralesional mineralization and pleural effusion negatively impacted accuracy.

Conclusions:

• The developed AI model demonstrates high accuracy and potential for automating pulmonary mass segmentation in canine CT imaging.
• This automated approach can significantly reduce the time and variability associated with manual segmentation in veterinary oncology.
• Further refinement may be needed to address challenges posed by complex lesion characteristics like mineralization or associated effusions.