Artificial intelligence–based detection of small and intermediate-to-large lymphocytes for cytologic diagnosis of canine lymphoma

Vollständiger Abstract

Worum geht es in dieser Arbeit?

Lymphoma is a common cancer in dogs, which often presents as generalized peripheral lymphadenopathy. Fine needle aspiration is frequently used to investigate the cause of peripheral lymphadenopathy, and artificial intelligence technology could potentially assist in cytologic interpretation. In this study, YOLOv11, an open-source object detection algorithm, was evaluated for cell-level identification in canine lymph node cytology to diagnose lymphoma. Cytologic images were captured using 2 smartphones from 11 non-lymphoma and 34 intermediate-to-large B-cell lymphoma peripheral lymph node aspirates in dogs, confirmed by cytologic interpretation plus either flow cytometry or polymerase chain reaction for antigen receptor rearrangements. A total of 25,761 intact cells were annotated across 680 images. Models were trained and validated under 5 cross-device configurations, comparing a 4-label approach (small, intermediate, and large lymphocytes and neutrophils) and a 3-label approach that merged intermediate and large lymphocytes. At the cell level, combining intermediate and large lymphocytes improved cell-level classification. In the optimized configuration, the model achieved an average precision of 87.60%, an average recall of 84.71%, a mean average precision 50 (mAP50) of 89.33%, and an F1 score of 86.02% on the independent test set. At the aspirate level, the model achieved near-perfect performance and correctly inferred diagnoses in all test aspirates by assessing the proportions of intermediate-to-large lymphocytes. These findings demonstrate the potential feasibility of smartphone-based deep learning assistance for veterinary cytology and highlight the importance of prospective, workflow-integrated validation, including external validation, across a broader range of lymph node diseases.