Prediction of canine lymphomas’ phenotype in cytological samples using a deep learning

Convolutional neural networks (CNN) are nowadays considered as the state-of-the-art image analysis and classification technique. In human medicine deep learning has achieved promising results in several different disciplines, but no reports about the possible applications in veterinary diagnostic pathology are currently available. In cytological classification of canine lymphoma, the possibility to define phenotype based on neoplastic lymphocytes morphology is still debated. The aim of the present study was to test the possibility to predict the immunophenotype of lymphoma, from cytological slides using deep learning. Thirty-one lymphoma cases (20 B-cell and 11 T-Cell) with flow cytometry-confirmed immunophenotype were retrospectively selected. Digital whole-slide images were acquired. Photographs of a variable number of individual cells (minimum 250) were obtained for each case. Images were than saved into different folders according to the immunophenotype. The images were randomly divided into a training, a validation and a test set using a 70%, 15%, 15% scheme. A pre-trained CNN (Inception V3) was retrained and fine-tuned on our database until an 80% accuracy in the test set was achieved. Thereafter, the diagnostic accuracy of the fine-tuned CNN was tested using a leave-one-out cross-validation. The predicted immunophenotype for each case was assigned calculating the modal predicted immunophenotype of the images. The immunophenotype was correctly predicted in 30/31 cases and the area under the curve was 0.96 (95% confidence interval (CI): 0.82-1), the sensitivity was 100% (CI: 82.35-100.00), and the specificity was 91.67% (CI: 61.52-99.79). Deep learning could potentially be used to predict the immunophenotype of lymphomas from cytological slides.