Lymphoma is one of the most common malignancies in dogs, and its diagnosis typically begins with fine needle aspiration of an enlarged peripheral lymph node. Cytology, the microscopic examination of cells harvested in this minimally invasive way, is fast and generally reliable, but it is not always definitive. Confirming a diagnosis of lymphoma and characterizing it as B-cell or T-cell disease usually requires histopathology with immunophenotyping through PCR for antigen receptor rearrangement, flow cytometry, or immunohistochemistry. Those ancillary tests add cost, time, and logistical complexity, and they are not uniformly available across veterinary practices. Because microscopic evaluation demands trained pathologists, high-quality optics, and consistent staining reagents, diagnostic delays can affect clinical decisions and outcomes for canine patients.

The Florida team asked whether a convolutional neural network, the class of deep learning architecture that has transformed medical image classification, could shoulder part of this diagnostic burden. They selected ResNet-50, a fifty-layer residual network pretrained on the ImageNet database, and fine-tuned it through transfer learning on cytological photomicrographs captured at 1,000-fold magnification with immersion oil. The researchers assembled their dataset from cases submitted between 2017 and 2024 to the University of Florida Small Animal Hospital, drawing on the archive of the Clinical Pathology Service and cross-checking electronic medical records to verify each diagnosis.

Strict inclusion and exclusion criteria shaped the final cohort. Lymphoma cases required a cytological diagnosis of intermediate or large-cell lymphoma confirmed by PARR or flow cytometry, while reactive lymphoid hyperplasia cases required a predominance of small mature lymphocytes with plasma cells and no concern for a neoplastic population. Three board-certified clinical pathologists independently reviewed every selected slide to ensure consensus. The resulting dataset comprised 260 lymph node slides from 184 dogs, from which 2,600 images were captured: 1,070 images of B-cell lymphoma from 72 dogs, 490 images of T-cell lymphoma from 35 dogs, and 1,040 images of reactive lymphoid hyperplasia from 77 dogs. Most slides were stained with Wright-Giemsa, with a small number stained with Diff-Quick to reflect real-world staining variation.

A critical design decision was that the dataset was partitioned at the patient level rather than the image level. This meant that no images from a dog in the test set ever appeared in the training or validation pools, preventing information leakage from inflating performance. Approximately 20 percent of patients were held out for final testing, while the remaining 80 percent were used for 10-fold cross-validation, with hyperparameter tuning and early stopping applied within each fold. Images were standardized to 1920 by 1080 pixels, normalized using ImageNet channel statistics computed without reference to the validation or test data, and augmented on the fly with random horizontal and vertical flips during training.

The first task was binary classification: lymphoma versus reactive lymphoid hyperplasia. Across ten independently trained models, test accuracy ranged from 85.4 percent to 95.4 percent, with a mean of 90.3 percent, and the area under the receiver operating characteristic curve ranged from 0.943 to 0.988. The best-performing model achieved an accuracy of 0.954 with a Matthews correlation coefficient of 0.904, sensitivity of 0.950, specificity of 0.958, and an F1 score of 0.957. The authors note that this level of sensitivity matches or slightly exceeds the roughly 93 percent sensitivity reported for clinical pathologists distinguishing lymphoma from non-lymphoma in cytology, and it clears the 94 percent threshold proposed for decision-rule validation in this domain. Training, validation, and test performance remained closely aligned across folds, suggesting minimal overfitting.

The second task was far harder: distinguishing B-cell from T-cell lymphoma using cytological appearance alone. This is notoriously difficult even for experienced pathologists, because both immunophenotypes often present as morphologically similar populations of large or intermediate lymphocytes, with only subtle clues such as occasional uropods in some T-cell cases. Unsurprisingly, the models performed more modestly here, with test accuracies between 66.2 percent and 79.8 percent, a mean of 74.3 percent, and area under the curve values between 0.653 and 0.748. The representative model achieved an accuracy of 0.743, though confidence intervals were wide given the limited number of T-cell patients. The authors point out that some discordance may stem from the reference standards themselves, since PARR and flow cytometry do not always agree, especially in samples with low tumor burden.

The study also carries important caveats that the authors address directly. The entire dataset came from a single veterinary teaching hospital, raising questions about geographic and equipment-related bias. Images were manually captured from representative fields chosen by pathologists rather than from whole-slide imaging, confirming high image quality but simplifying the task relative to real diagnostic workflows. Cases of small-cell lymphoma, which account for fewer than two percent of canine lymphomas and are extremely difficult to separate from reactive hyperplasia on cytology alone, were deliberately excluded, meaning the model’s scope is limited to the most cytologically distinct presentations. Class imbalance and the predominance of one imaging system may also have influenced performance, and the researchers emphasize that robust external validation on prospectively collected, multi-institutional data is essential before any clinical deployment.

Transparency was a stated priority throughout. Unlike conventional practice, where a hold-out test set is evaluated once, the team reported results from all ten cross-validation folds and identified the best-performing model, providing a detailed suite of metrics including confidence intervals derived from patient-level bootstrapping with 1,000 iterations. The source code for preprocessing, training, and evaluation will be released publicly on GitHub, and the anonymized image dataset is available from the corresponding author upon reasonable request under a data use agreement. The authors argue that this openness stands in contrast to the proprietary opacity that often surrounds commercial AI diagnostic tools, such as the cytology-triage platforms already entering veterinary practice.

Looking forward, the researchers frame their models as triage and decision-support tools rather than replacements for pathologists, insisting that a board-certified clinical pathologist must retain final diagnostic responsibility and interpret model outputs alongside clinical history and ancillary tests. Future work will explore whole-slide imaging pipelines that could mirror the way pathologists actually examine slides, explainable AI techniques such as Gradient-weighted Class Activation Mapping to reveal which cellular features drive predictions, other CNN architectures, and multi-institutional datasets spanning diverse stains, scanners, and patient populations. The moderate success of B-cell versus T-cell classification, despite the well-known morphological overlap between these immunophenotypes, hints that deep networks may be extracting genuinely subtle cytomorphological patterns. For now, the study establishes a careful, reproducible performance baseline and an explicit methodological roadmap for bringing AI-assisted cytology from the research bench toward the veterinary clinic, where faster answers could mean earlier treatment for dogs with lymphoma.

Subject of Research: Deep learning classification of canine lymphoma in lymph node cytology images

Article Title: Deep learning-assisted cytological image analysis for canine lymphoma

Article References: Deep learning-assisted cytological image analysis for canine lymphoma. (n.d.). https://doi.org/10.1186/s44356-026-00056-5

Image Credits: AI Generated

DOI: 10.1186/s44356-026-00056-5

Keywords: canine lymphoma, deep learning, cytology, ResNet-50, veterinary pathology, convolutional neural networks, B-cell lymphoma, T-cell lymphoma, reactive lymphoid hyperplasia, image classification, veterinary oncology, artificial intelligence