Lymphoma (also called lymphosarcoma or LSA) is a prevalent malignancy in dogs with cytological evaluation being an efficient and generally reliable method for initial clinical assessment of suspected LSA cases. However, definitive characterization requires histopathology supported by immunophenotyping, which may not be readily available in all clinical settings. This study investigated the potential of using deep learning to automate cytological diagnosis, specifically assessing the capability of a convolutional neural network (CNN), ResNet-50, to differentiate LSA from reactive lymphoid hyperplasia (RLH) and to further classify LSA cases into B-cell and T-cell phenotypes. Cytological images (1,000× magnification) were captured from 260 lymph node slides obtained from 184 dogs, yielding 2,600 images comprising 1,070 B-cell LSA images (72 dogs), 490 T-cell LSA images (35 dogs), and 1,040 RLH images (77 dogs). The dataset was partitioned at the patient level into training, validation, and test subsets, and a 10-fold cross-validation approach was applied. The ResNet-50 CNN differentiated LSA from RLH with a test accuracy of 85.4% − 95.4% and an area under the receiver operating characteristic curve (AUC-ROC) of 0.943–0.988 across ten models with low variability. The model classifying B-cell versus T-cell LSA revealed a test accuracy of 66.2% − 79.8% and an AUC-ROC of 0.653–0.748 across ten models. These results demonstrate a proof-of-concept study exploring the potential of deep learning for the cytological evaluation of canine LSAs, including B-cell and T-cell phenotyping. This study also provides an explicit methodology of deep learning applications in veterinary pathology images, enhancing further development.
Shum et al. (Sun,) studied this question.