Deep learning models for image classification of lymphoma: a pilot study in canine

Area of Science:

• Veterinary Medicine
• Artificial Intelligence
• Image Analysis

Background:

• Canine lymphoma diagnosis can be challenging, often requiring differentiation from reactive lymphoid hyperplasia (RLH).
• Fine needle aspiration (FNA) is a common diagnostic tool, but image interpretation requires expertise.

Purpose of the Study:

• To develop and evaluate deep learning models for distinguishing canine lymphoma from RLH using FNA images.
• To compare the performance of Vision Transformer (ViT) and Inception-v3 models, including ensemble approaches.

Main Methods:

• Four deep learning models were developed: Vision Transformer (ViT), Inception-v3, and two ensemble models (MEAN and MAX) combining ViT and Inception-v3.
• A dataset of 2,290 canine lymphoma FNA images and 871 RLH FNA images was utilized.
• Three distinct training and testing datasets were prepared for robust model evaluation.

Main Results:

• Inception-v3 and the two ensemble models achieved high performance metrics (>80% accuracy, recall, and AUC) across all datasets.
• The ViT model showed high precision (>0.85) but did not reach comparable performance in other metrics.
• Deep learning models demonstrated significant potential in classifying canine lymphoma from FNA images.

Conclusions:

• Deep learning, particularly ensemble methods and Inception-v3, shows strong potential for accurate canine lymphoma diagnosis from FNA images.
• These AI-driven approaches can aid veterinarians in differentiating lymphoma from RLH, improving diagnostic efficiency.
• Further research can explore larger datasets and diverse model architectures to enhance diagnostic capabilities.