Applicability of explainable artificial intelligence in cytopathological evaluation of canine lymphoma

Applicability of explainable artificial intelligence in cytopathological evaluation of canine lymphoma | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Research Article Applicability of explainable artificial intelligence in cytopathological evaluation of canine lymphoma Chanel Shum, Donghee Lee, Karthik Karnam, Matthew Vardas, Samantha Masessa, and 6 more This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-9225846/v1 This work is licensed under a CC BY 4.0 License Status: Under Review Version 1 posted 6 You are reading this latest preprint version Abstract Background : Explainable artificial intelligence (AI) is a valuable tool for enhancing transparency in AI-driven image classification applications. Its practical utility and clinical evaluation remain limited in veterinary oncology, particularly within the field of diagnostic imaging. The aim of this study was to develop explainable image-based classification models using cytological data from canine lymphoma and reactive lymphoid hyperplasia (RLH). Methods : A total of 5,378 images were generated from cytological samples collected from 182 dogs, including 2,284 images of lymphoma (105 dogs) and 3,094 images of RLH (77 dogs). The dataset was stratified by patient and divided into training, validation, and test subsets, with model robustness evaluated through a 10-fold cross-validation protocol. Binary classification models were established using the ResNet-50 architecture. Model interpretability in a clinical context was assessed using the Gradient-weighted Class Activation Mapping (Grad-CAM) technique. Results: The model for classifying lymphoma and RLH images achieved a test accuracy of 0.945 (95% CI: 0.929 - 0.958), a weighted precision of 0.945 (95% CI: 0.903 - 0.939), a weighted recall of 0.945 (95% CI: 0.899 - 0.969), a weighted F1-score of 0.945 (95% CI: 0.910 - 0.943), an area under the receiver operating characteristic curve of 0.979 (95% CI: 0.957 - 0.978), and an area under the precision-recall curve of 0.981 (95% CI: 0.960 - 0.985). At the patient level, 18 of 20 dogs with lymphoma were correctly classified (true positives), and all 16 dogs with RLH were correctly classified (true negatives), corresponding to a sensitivity of 0.900 (95% CI: 0.699 - 0.972), specificity of 1.000 (95% CI: 0.806 - 1.000), positive predictive value of 1.000 (95% CI: 0.824 - 1.000), and negative predictive value of 0.889 (95% CI: 0.672 - 0.969). Grad-CAM provided visual explanations of the model predictions by highlighting image regions that reflect cellular morphology pertinent to lymphoma diagnostics. Conclusions : Our findings suggest that explainable AI is applicable to cytological analysis of canine lymphoma, offering pathological insights through visual explanation of model predictions. Artificial intelligence Canine lymphoma Convolutional neural networks Cytology Deep learning Explainable AI Grad-CAM ResNet-50 Figures Figure 1 Figure 2 Figure 3 Figure 4 Figure 5 Figure 6 Figure 7 Background Artificial intelligence (AI) is a powerful tool for revolutionizing veterinary biomedical research, establishing a solid foundation for precision-driven digital health management in animal patients, especially companion animals [1, 2]. AI applications span disciplinary boundaries and may address gaps in veterinary knowledge [1, 3]. In particular, health management for companion animals often mirrors human medical systems, creating a bidirectional framework of reciprocal benefit. This convergence enables the development of clinically robust translational models that circumvent the ethical and biological limitations inherent in traditional laboratory animal research [4]. In this regard, the integration of AI and machine learning into veterinary oncology exemplifies the shift, offering scalable, data-intensive platforms that enhance diagnostic precision and accelerate therapeutic discovery. Yet, despite its transformative potential, the current stage of AI integration in veterinary medicine remains nascent. Successful clinical applications require rigorous validation and interdisciplinary collaboration to ensure reliability, interpretability, and clinical utility. Lymphoma is the most prevalent hematopoietic malignancy in domestic dogs, comprising approximately 80% of all canine hematopoietic neoplasms [5, 6]. Reactive lymphoid hyperplasia (RLH), a non-neoplastic lymph node enlargement characterized by a heterogeneous lymphoid population, represents the primary cytologic differential diagnosis for lymphoma [7]. Initial diagnosis and treatment often begin in local veterinary clinics through physical examination and cytological evaluation of fine-needle aspirates. While histopathological examination offers the gold-standard for definite diagnosis [5, 8], cytology offers a cost-effective and practical alternative, especially for cases where financial constraints or limited access to advanced diagnostics pose challenges for pet owners and clinics [8-11]. This approach is widely adopted in routine practice and often yields outcomes comparable to those achieved in tertiary veterinary centers. Fine-needle aspirate cytology is sufficient to establish a diagnosis in most dogs with typical multicentric lymphoma [12]. However, excisional or core biopsy is recommended when cytological findings are equivocal, when low-grade lymphoma is suspected, or when full histological classification is required. Despite these guidelines, diagnostic and treatment decisions in local settings often rely on the anecdotal experience of general practitioners, particularly when input from clinical pathologists is unavailable. Standardized guidelines for cytological assessment for lymphoma that support more consistent evaluations are available to veterinary practitioners and medical oncologists [7, 13]. These guidelines provide a framework for morphometric analysis of cytological images; however, they remain largely dependent on semi-quantitative assessments performed by the human visual system and require professional training. Recent advances and research endeavors have demonstrated the application of deep learning techniques to canine lymphoma utilizing digital histopathological images [14-17]. Nevertheless, clinically reliable digital systems for automated cytological analysis have yet to be established. In our previous study, we employed the ResNet-50 deep learning framework to evaluate the capability of pre-trained convolutional neural networks (CNNs) in classifying cytological images of canine lymphoma and RLH [18]. Despite the strong performance of the classification models, the study provided limited insights into model interpretability. Here, we present new approaches to refine the ResNet-50 models and to establish explainable AI frameworks that enhance interpretability and robustness in the cytopathological analysis of canine lymphoma. Methods Datasets We utilized datasets of cytological images of canine lymphoma and RLH established in our previous study [18]. In the current study, these images were re-processed for data augmentation and mining to enhance their pathological relevance. Specifically, the 2,600 original images, prepared at 1920 × 1080 pixels, were divided into four sub-images of 960 × 540 pixels each, resulting in a total of 10,400 images. Color representation was standardized and images were normalized using ImageNet mean and standard deviation values of the RGB channels, as previously described [18]. Subsequently, all individual images underwent thorough review by three investigators: two clinical pathology residents (MV and CS) and one veterinary student (SM), all of whom were adequately trained in cytology image analysis by a board-certified clinical pathologist (CDS). During evaluation, we applied specific criteria to ensure the quality of images, focusing on those representing sufficient cellular components for pathological assessment and lymphoma diagnosis. The inclusion criteria required each image to contain more than 10 viable nucleated cells and more than seven viable lymphocytes. For images representing diagnosed lymphoma, only those containing at least 70% large lymphocytes among the total lymphocytes were included. Images of poor quality, such as those that were out of focus or had excessively high cell density preventing individual cell distinction, were excluded. This quality control review produced image datasets with adequate components per image, suitable for training purposes to differentiate between lymphoma and RLH. The final image dataset consisted of 5,378 images, including 2,284 images of lymphoma (105 dogs) and 3,094 images of RLH (77 dogs). Representative images for each group are provided ( Figs. 1 and 2 ). Programming environment All experiments were conducted in a Python (version 3.10.12) environment on a Red Hat Enterprise Linux (RHEL) 8 operating system. The deep learning models were implemented, trained, and evaluated using PyTorch (version 2.5.1) and its accompanying torchvision library (version 0.20.1) for model access and image transformations. GPU acceleration was provided by NVIDIA A100 GPUs (80 GB VRAM each), utilizing the CUDA Toolkit (version 11.8). Key supporting libraries included scikit-learn (version 1.4.0) for calculating performance metrics, NumPy (version 1.26.4) for numerical operations, and Pillow (version 11.0.0) for image loading and initial manipulation. All computational tasks were performed on the University of Florida's research computing cluster, HiPerGator, equipped with an AMD EPYC 75F3 Milan CPU (64 cores) and 500 GB of RAM. Model development Data splitting and model training were performed as previously described [18]. Briefly, digital images that passed quality control were divided at the dog-patient level into training and hold-out test sets in an 8:2 ratio. Partitioning was done by randomly shuffling patients using numpy.random.permutation (NumPy version 1.26.4) with a fixed random seed. Approximately 20% of patients from each disease group were allocated to the test set, with the remaining 80% used for training and validation. Table 1 summarizes the dataset composition, including the numbers of dog patients and images. Additional data augmentation including random vertical and horizontal flipping was conducted using torchvision.transforms.RandomVerticalFlip(p=0.5) and torchvision.transforms.RandomHorizontalFlip(p=0.5) (Torchvision version 0.20.1). Random 90° rotation and color jitter (brightness ± 20%, contrast ± 20%, saturation ± 10%, hue ± 0.02) were also applied. Transfer learning was then performed on the training/validation datasets using the ResNet-50 CNN architecture initialized with weights pretrained on ImageNet-1K (ILSVRC2012) [19]. The final fully connected layer was replaced with a single-unit linear layer to enable binary classification, followed by sigmoid activation to generate probability outputs. Hyperparameter tuning was conducted via 10-fold cross-validation. For this process, the datasets excluding the test set were randomly divided at the dog-patient level into ten distinct groups in a 9:1 ratio for training and validation subsets. Training and validation were repeated across all ten folds. Early stopping was employed by monitoring validation loss, with training halted if no improvement was observed for four consecutive epochs (patience = 4), up to a maximum of 50 epochs. The model was trained using the Adam optimizer (learning rate = 1e-4, no scheduler, no weight decay) with a batch size of 2. Binary cross-entropy loss was used without class weighting. All layers were kept unfrozen during training. After each fold, accuracy and loss were computed and recorded for both training and validation sets. Model evaluation After hyperparameter optimization through 10-fold cross-validation on the training/validation datasets, we evaluated the model on the hold-out test set (382 lymphoma images from 20 dogs; 601 RLH images from 16 dogs). Performance was assessed at two levels: image- and patient-level. The performance metrics included accuracy, weighted average precision, weighted average recall, weighted average F1-score, and the area under the receiver operating characteristic curve (AUC-ROC), and the area under the precision-recall curve (AUC-PR). Considering the imbalanced distribution of the datasets among the classification groups, model performance for each fold was evaluated using this comprehensive set of metrics, alongside training and validation curves, as previously described [20, 21]. 95% confidence intervals (CIs) were estimated using the Wilson score method. Image-level predictions were generated using a classification threshold of 0.5, and a majority-vote rule was applied to determine the classification of each dog at the patient level based on the predominant image-level predictions. Gradient-weighted Class Activation Mapping (Grad-CAM) Grad-CAM was employed to provide visual explanations from our developed model [22]. Visualizations were generated using the pytorch-grad-cam implementation applied to the last convolutional block of ResNet-50. For each image, class-specific gradients were back-propagated to this layer, gradients were globally averaged to weight feature maps, the weighted sum was passed through ReLU, normalized, upsampled to the input resolution, and overlaid on the original RGB image. Grad-CAMs were produced during evaluation for representative cases and exported in batch for the entire test set, and outputs were saved to the results directory while preserving class and patient subfolders with a manifest CSV of file paths and labels. Results ResNet-50-based models show robust and consistent learning for binary classification of canine lymphoma and RLH cytological images First, we developed ten deep learning models using 1,902 lymphoma images from 85 dogs and 2,493 RLH images from 61 dogs through 10-fold cross-validation. We found that the models exhibited robust and consistent learning behavior across folds. On average, training accuracy increased from 0.826 to 0.972, while validation accuracy improved from 0.887 to 0.975 ( Figs. 3A, B ). By the final optimized epoch, training accuracy remained consistently high across all folds, reaching between 0.951 and 0.970, indicating strong model fit to the training data. Validation accuracy improved steadily with each epoch, with most folds achieving final values between 0.907 and 0.986, reflecting robust generalization. Training loss decreased from 0.390 to 0.074, and validation loss followed a comparable downward trend, albeit with slightly higher values and greater variability than training loss ( Figs. 3C, D ). Training loss declined consistently across epochs in all folds, often reaching values below 0.10. Validation loss also decreased in most folds, though some exhibited plateaus or fluctuations. Collectively, the training and validation curves across all folds suggest effective learning, convergence, and minimal overfitting. Complete metrics of training and validation performance over epochs using 10-fold cross-validation are provided in Supplementary Table S1 . Evaluation of test performance demonstrates model suitability and clinically relevant insights into diagnostic utility. With optimized hyperparameters, we assessed a final model performance using a hold-out test dataset comprising 382 lymphoma images from 20 dogs and 601 RLH images from 16 dogs. The performance metrics included accuracy, weighted precision, weighted recall, weighted F1-score, AUC-ROC, and AUC-PR. First, at the image level, with patient-level dataset splitting, the model demonstrated a test accuracy of 0.945 (95% CI: 0.929 - 0.958), a weighted precision of 0.945 (95% CI: 0.903 - 0.939), a weighted recall of 0.945 (95% CI: 0.899 - 0.969), a weighted F1-score of 0.945 (95% CI: 0.910 - 0.943), an AUC-ROC of 0.979 (95% CI: 0.957 - 0.978), and an AUC-PR of 0.981 (95% CI: 0.960 - 0.985) ( Fig. 4 ). Then, to assess clinically meaningful diagnostic performance, we evaluated the model at the patient level using a majority vote rule, classifying each dog based on the predominant image-level predictions. Under this rule, 18 of 20 lymphoma dogs were correctly classified (true positives), and all 16 RLH dogs were correctly classified (true negatives). This corresponded to a sensitivity of 0.900 (95% CI: 0.699 - 0.972), specificity of 1.000 (95% CI: 0.806 - 1.000), positive predictive value of 1.000 (95% CI: 0.824 - 1.000), and negative predictive value of 0.889 (95% CI: 0.672 - 0.969). These findings indicate that our ResNet-50-based model is well-suited for binary classification of canine lymphoma and RLH cytological images. Also, patient-level performance evaluation provides clinically relevant insights into diagnostic utility. Grad-CAM supports morphological identification of malignant lymphocytes distinct from benign cells To address model explainability, we deployed Grad-CAM, which visualizes the decision-making features within the cytological images. We generated Grad-CAM heatmaps for all test set images from both groups. The heatmaps highlighted regions enriched with large lymphocytes in lymphoma images ( Fig. 5 ) and regions enriched with small mature lymphocytes and rare plasma cells in RLH images ( Fig. 6 ). Despite considerable morphological diversity, cellular composition, and color variation across groups, these Grad-CAM findings demonstrate that the model extracted deterministic features differentiating lymphoma from RLH, concordant with cellular characteristics used for diagnostic identification. As described above, test performance was evaluated at both the image and patient level. To further enhance the model interpretability, we calculated the proportions of correctly classified images for each dog patient. Among lymphoma cases, 18 of 20 dogs had more than 66.7% of their images correctly classified; specifically, 16 dogs exceeded 97% and two dogs had 84.6% and 66.7%, respectively ( Fig. 7A; Supplementary Table S2 ). Two lymphoma cases (Patients #19 and #20) demonstrated markedly lower correct classification rates of 17.4% and 0%. For RLH cases, all 16 dogs had at least 70% of their images correctly identified, representing true negatives ( Fig. 7B; Supplementary Table S3 ). To investigate the two lymphoma cases with high rates of misclassified images (i.e., false negatives), we examined the Grad-CAM results. In Patient #19, cytological evaluation revealed an abnormal dimorphic lymphocyte population characterized by small mature lymphocytes and an expanded population of large immature lymphocytes, with an absence of intermediate-size lymphocytes, findings that are morphologically diagnostic of lymphoma ( Fig. 7C ). However, only 4 of 23 images were correctly predicted, with the Grad-CAM highlighting regions containing large lymphocytes. The other 19 images were misclassified as RLH, with the Grad-CAM highlighting regions containing small or normal-sized lymphocytes. This pattern suggests that our model distinguishes lymphoma and RLH based on differences in cell size between large and small lymphocytes. In Patient #20, all four images were misclassified despite meeting all pre-processing criteria; they showed relatively low cellularity ( Fig. 7D ). Grad-CAM highlighted areas containing small lymphocytes, suggesting that samples with reduced cellularity and heterogeneous morphology across different areas may also contribute to false-negative predictions. Collectively, our observations suggest that Grad-CAM technique may aid model interpretability by highlighting image regions, although these visualizations should be interpreted as coarse regional associations rather than cell-specific explanations. Discussion This study shows that a ResNet-50 deep learning architecture facilitates reliable model development for AI-assisted identification of canine lymphoma using cytological images. It also provides a useful approach for integrating explainable AI tools to enhance model explainability in the context of cytological characterization. Our work presents a computational workflow that mirrors key aspects of the morphological differentiation of malignant lymphoid cells from benign conditions in canine specimens, thereby supporting technological advancement in veterinary cytopathology. ResNet-50 is a widely utilized deep learning architecture in computer vision [23]. Its effectiveness, along with strong support from both the scientific community and industry, has made pretrained ResNet-50 models a popular choice for transfer learning [24-27]. This widespread adoption offers practical advantages to developers and users, such as robust technical resources and guidance for mitigating bias during model development and subsequent refinement [28]. Although AI-driven image classification has become popular, deep learning applications in veterinary cytology are still limited, at least in part, due to restricted technical accessibility in the veterinary field. Tasks require advanced computational expertise and high-performance hardware. However, these barriers are rapidly receding with ongoing advances in computing power and development tools. Cytological examinations rely on holistic morphological assessment by expert clinical pathologists. Consequently, employing ResNet-50 for cytological image analysis represents a viable strategy, with reasonable prospects for success, provided the images contain sufficient morphological features for the model to learn. While this potential has been explored in human clinical cytopathology [29-31], we recently extended it to veterinary medicine. We previously developed prototype models to differentiate canine lymphoma from RLH, and to distinguish B-cell from T-cell lymphoma using a dataset of 2,600 images captured at 1920 × 1080 pixels [18]. Despite rigorous technical validation and comprehensive performance evaluation, this prior work lacked model explainability, possibly because cell-level segmentation and annotation were not performed. In the current study, we continued to employ weakly supervised, image-level annotation, but supplemented it with rigorous human intervention to enhance model interpretability. Specifically, a dataset of 5,378 images, pre-processed at 960 × 540 pixels, underwent additional quality control by clinical pathologists to ensure that each image contained a sufficient number of viable nucleated cells and large lymphocytes. The assessment of individual images was not definitive for lymphoma diagnosis; in practice, pathologists review entire slides in conjunction with clinical history. Accordingly, our deep learning models do not yet fully replicate the comprehensive diagnostic process performed by clinicians. The goal of this study was not to replace pathologists but to evaluate whether ResNet-50 deep learning architecture has practical utility for cytological image analysis in veterinary oncology [32, 33]. Our results suggest that transfer learning with pretrained ResNet-50 models is feasible for canine lymphoma cytology, supporting both the versatility and broader applicability of this approach. By emphasizing transparency in model development, this work addresses concerns related to unclear or opaque procedures that can undermine reliability and trust in AI models in the veterinary domain [33-36]. The integration of explainable AI tools further strengthens this effort: Grad-CAM visualizations provided evidence of differential morphological features captured by the models between lymphoma and RLH cases. Intriguingly, similar to human visual assessment, the models appeared to focus on image regions that reflect features typically used to differentiate benign from malignant conditions. However, reliance on a single explainability method such as Grad-CAM may raise concerns regarding faithfulness and completeness, and alternative explainable AI techniques are available [37]. Thus, future studies should include comparative evaluations of multiple explainability approaches across different model architectures. The misclassified lymphoma cases are unlikely random errors but reflect systematic vulnerabilities of the model. Plasma cell-rich or plasmacytoid variants represent a well-recognized diagnostic challenge, even for experienced pathologists [9]. Similarly, low-cellularity samples, which can result from hemodilution or suboptimal aspiration technique, are a common pre-analytical issue in cytology and may be more frequent among general practitioners. At the patient level, the model exhibited a 10% false-negative rate (2/20 lymphoma cases), with a negative predictive value of 0.889, indicating that approximately 1 in 9 cases predicted as “negative” may in fact represent lymphoma. Clinically, such false negatives could delay diagnosis and treatment, underscoring the need for cautious interpretation of negative results, particularly in cases with atypical morphology or low cellularity. Potential mitigation strategies include implementing a minimum confidence threshold for automated predictions or incorporating a pre-screening cellularity check to flag low-cellularity images for manual review, thereby improving safety in real-world deployment. A key limitation of this study is that the model classifies only lymphoma versus RLH. In clinical practice, lymph node enlargement may result from a broad range of conditions, including poorly cohesive metastatic carcinoma, melanoma, mast cell tumor, histiocytic sarcoma, granulomatous or suppurative lymphadenitis, and others. As a closed-set binary classifier, the model necessarily assigns every image to one of the two predefined categories. A lymph node containing metastatic neoplasia that is misclassified as RLH would represent a clinically significant false negative that the current evaluation framework cannot detect. Future work should explore multi-class classification models or open-set recognition approaches to better reflect real-world diagnostic complexity. Beyond technical and clinical performance, it is essential to address the professional skepticism and heightened caution prevalent within the veterinary community [33, 34, 38]. The primary barrier to AI adoption is often not the technology itself, but a lack of transparency and rigor during development. From a pathologist’s perspective, AI still requires significant human oversight, including case review, quality control, data collection, in-depth cellular segmentation, and micrometer-resolution annotation. Despite the necessary involvement of human experts, scientific publications and commercial products often describe AI capabilities in ways that can seem overstated or unclear. This disparity creates a lack of accountability that hinders clinical trust. By prioritizing rigorous technical validation and model explainability, our study provides the transparency necessary to mitigate these concerns. Although the FDA and other bodies have yet to establish comprehensive regulatory frameworks to keep pace with these innovations [36], technical transparency and regulatory guidance must advance in tandem to ensure the safe and reliable integration of AI into veterinary medicine. In conclusion, our study demonstrates a feasible and transparent approach for developing AI-assisted cytological classification in veterinary medicine. By moving beyond “black-box” methodologies and incorporating human-verified datasets and Grad-CAM visualizations, our fundings suggest that AI systems can be developed with rigor appropriate for preliminary clinical evaluation. To further refine these models and assess the clinical reliability, multi-institutional studies with large research consortia are needed. While these models do not replace the holistic expertise of a pathologist, they offer a powerful digitized platform for enhancing diagnostic workflows. As technical and regulatory guidance continues to advance in tandem, this transparent developmental model provides a template for the responsible and trustworthy implementation of AI in veterinary practice. Abbreviations AI: artificial intelligence AUC-PR: area under the precision-recall curve AUC-ROC: area under the receiver operating characteristic curve CI: confidence interval CNN: convolutional neural networks CPU: central processing unit CSV: comma-separated values CUDA: compute unified device architecture GPU: graphics processing unit Grad-CAM: gradient-weighted class activation mapping IACUC: institutional animal care and use committee RAM: random access memory RGB: red-green-blue RHEL: Red Hat Enterprise Linux RLH: reactive lymphoid hyperplasia VRAM: video random access memory Declarations Ethics approval and consent to participate This retrospective study analyzed existing, de-identified cytology images from client-owned dogs. Veterinary clinical samples were collected under standard clinical care with owner consent in accordance with institutional policy, or institutional review determined that consent was not required for de-identified retrospective data. No prospective enrollment, procedures, or sample collection from live animals were performed; thus, formal Institutional Animal Care and Use Committee (IACUC) review was not required. This determination is consistent with PHS/OLAW guidance on the scope of IACUC oversight. Consent for publication Not applicable. Availability of data and materials The source code used for image preprocessing, model training, and evaluation in this study will be made publicly available on GitHub at [Placeholder for URL – e.g., to be provided upon acceptance and code upload]. The anonymized cytological image dataset supporting the findings of this study is available from the corresponding author (J.H.K and C.S.) upon reasonable request, subject to institutional guidelines and a data use agreement. Competing interests The authors declare that they have no competing interests. Funding The authors disclosed the following financial support for the research, authorship, and publication of this article: this work was partially supported by the CVM Clinical Pathology Residency program, the UF Strategic Funding Initiative, AI Initiative, AI Scholar Award, and Lisa Conti One Health Initiative through the Florida Veterinary Scholars Program (FVSP) at the University of Florida. This work was also supported by the Mid-Florida Golden Retriever Club. Authors' contributions JHK and CDS designed the study and protocols; CS, MV, SM, KH, SG, and CDS collected the samples, performed microscopic evaluations, and captured images; DL, KK, HJ, and KC performed data preprocessing and computational work; KH contributed to administrative support; the manuscript was written by CS, DL, KK, MV, JHK, and CDS with contribution from all other authors. Acknowledgements We would like to acknowledge the staff and technicians at the clinical pathology service for supporting sample collection and data curation at the University of Florida Small Animal Hospital. The authors also acknowledge computational resources and technical support provided by the HiPerGator Research Computing at the University of Florida. References Akinsulie OC, Idris I, Aliyu VA, Shahzad S, Banwo OG, Ogunleye SC, Olorunshola M, Okedoyin DO, Ugwu C, Oladapo IP et al : The potential application of artificial intelligence in veterinary clinical practice and biomedical research . Front Vet Sci 2024, 11 :1347550. Arshad MF, Ahmed F, Nonnis F, Tamponi C, Scala A, Varcasia A: Artificial intelligence and companion animals: Perspectives on digital healthcare for dogs, cats, and pet ownership . Res Vet Sci 2025, 193 :105776. Basran PS, Appleby RB: The unmet potential of artificial intelligence in veterinary medicine . Am J Vet Res 2022, 83 (5):385-392. Oh JH, Cho JY: Comparative oncology: overcoming human cancer through companion animal studies . Exp Mol Med 2023, 55 (4):725-734. Zandvliet M: Canine lymphoma: a review . Vet Q 2016, 36 (2):76-104. Vail D, Thamm D, Liptak J: Hematopoietic Tumors. In: Withrow and MacEwen's Small Animal Clinical Oncology. edn.; 2019: 688-772. Raskin RE, Meyer D: Canine and Feline Cytology—A Color Atlas and Interpretation Guide , 2nd edn. St. Louis, MO: Elsevier; 2009. Seelig DM, Avery AC, Ehrhart EJ, Linden MA: The Comparative Diagnostic Features of Canine and Human Lymphoma . Vet Sci 2016, 3 (2). Aresu L: Canine Lymphoma, More Than a Morphological Diagnosis: What We Have Learned about Diffuse Large B-Cell Lymphoma . Front Vet Sci 2016, 3 :77. Riondato F, Comazzi S: Flow Cytometry in the Diagnosis of Canine B-Cell Lymphoma . Front Vet Sci 2021, 8 :600986. Das DK: Contribution of Immunocytochemistry to the Diagnosis of Usual and Unusual Lymphoma Cases . J Cytol 2018, 35 (3):163-169. Childress MO, Avery A, Behling-Kelly E, Bennett P, Brockley L, Dickinson R, Hughes K, Kisseberth WC, Marconato L, Martini V et al : Diagnosis and Classification of Primary Nodal Lymphomas in Dogs: A Consensus of the Oncology-Pathology Working Group . Vet Comp Oncol 2025, 23 (3):331-345. Valenciano AC, Cowell RL: Cowell and Tyler’s Diagnostic Cytology and Hematology of the Dog and Cat , 5th edn. St. Louis, MO: Elsevier; 2020. Hubbard-Perez M, Luchian A, Milford C, Ressel L: Use of deep learning for the classification of hyperplastic lymph node and common subtypes of canine lymphomas: a preliminary study . Front Vet Sci 2023, 10 :1309877. Haghofer A, Fuchs-Baumgartinger A, Lipnik K, Klopfleisch R, Aubreville M, Scharinger J, Weissenböck H, Winkler SM, Bertram CA: Histological classification of canine and feline lymphoma using a modular approach based on deep learning and advanced image processing . Sci Rep 2023, 13 (1):19436. Ancheta K, Le Calvez S, Williams J: The digital revolution in veterinary pathology . J Comp Pathol 2024, 214 :19-31. Ancheta K, Psifidi A, Yale AD, Le Calvez S, Williams J: Deep-learning based morphological segmentation of canine diffuse large B-cell lymphoma . Front Vet Sci 2025, 12 :1656976. Shum C, Lee D, Jo H, Kim J, Heo H, Gilor S, Lanier C, Hall K, Dark M, Kim JH et al : Deep learning-assisted cytological image analysis for canine lymphoma . Veterinary Oncology 2026, 3 (1):5. Russakovsky O, Deng J, Su H, Krause J, Satheesh S, Ma S, Huang Z, Karpathy A, Khosla A, Bernstein M et al : ImageNet Large Scale Visual Recognition Challenge. Int J Comput Vis 2015, 115 (3):211–252. Maier-Hein L, Reinke A, Godau P, Tizabi MD, Buettner F, Christodoulou E, Glocker B, Isensee F, Kleesiek J, Kozubek M et al : Metrics reloaded: recommendations for image analysis validation . Nat Methods 2024, 21 (2):195-212. Gaudreault J-G, Branco P: Empirical analysis of performance assessment for imbalanced classification . Machine Learning 2024, 113 (8):5533-5575. Selvaraju RR, Cogswell M, Das A, Vedantam R, Parikh D, Batra D: Grad-CAM: Visual Explanations from Deep Networks via Gradient-Based Localization . In: 2017 IEEE International Conference on Computer Vision (ICCV). Venice, Italy; 2017: 618-626. He K, Zhang X, Ren S, Sun J: Deep Residual Learning for Image Recognition . In: 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR): 27-30 June 2016 2016 ; 2016: 770-778. Zaidi SSA, Ansari MS, Aslam A, Kanwal N, Asghar M, Lee B: A survey of modern deep learning based object detection models . Digital Signal Processing 2022, 126 :103514. Desanamukula VS, Teja TD, Rajitha P: An In-Depth Exploration of ResNet-50 and Transfer Learning in Plant Disease Diagnosis . In: 2024 International Conference on Inventive Computation Technologies (ICICT): 24-26 April 2024 2024 ; 2024: 614-621. Alzubaidi L, Zhang J, Humaidi AJ, Al-Dujaili A, Duan Y, Al-Shamma O, Santamaría J, Fadhel MA, Al-Amidie M, Farhan L: Review of deep learning: concepts, CNN architectures, challenges, applications, future directions . J Big Data 2021, 8 (1):53. Jena B, Nayak GK, Saxena S: Convolutional neural network and its pretrained models for image classification and object detection: A survey . Concurr Comput Pract Exp 2022, 34 (6):e6767. Huang Y, Lin L, Cheng P, Lyu J, Tam R, Tang X: Identifying the Key Components in ResNet-50 for Diabetic Retinopathy Grading from Fundus Images: A Systematic Investigation . Diagnostics (Basel) 2023, 13 (10). Su F, Sun Y, Hu Y, Yuan P, Wang X, Wang Q, Li J, Ji J-F: Development and validation of a deep learning system for ascites cytopathology interpretation . Gastric Cancer 2020, 23 (6):1041-1050. Kaur H, Sharma R, Kaur J: Comparison of deep transfer learning models for classification of cervical cancer from pap smear images . Sci Rep 2025, 15 (1):3945. Rodríguez M, Córdova C, Benjumeda I, San Martín S: Automated Cervical Cancer Screening Using Single-Cell Segmentation and Deep Learning: Enhanced Performance with Liquid-Based Cytology . Computation 2024, 12 (12):232. Pacholec C, Flatland B, Xie H, Zimmerman K: Harnessing artificial intelligence for enhanced veterinary diagnostics: A look to quality assurance, Part II External validation . Vet Clin Pathol 2025. Sobkowich KE: Demystifying artificial intelligence for veterinary professionals: practical applications and future potential . Am J Vet Res 2025, 86 (S1):S6-S15. Cazer CL, Basran P, Ivanek-Miojevic R: From bark to bytes: artificial intelligence transforming veterinary medicine . Am J Vet Res 2025, 86 (S1):S4-S5. Cohen EB, Gordon IK: First, do no harm. Ethical and legal issues of artificial intelligence and machine learning in veterinary radiology and radiation oncology . Vet Radiol Ultrasound 2022, 63 Suppl 1 (Suppl 1):840-850. Duggirala HJ, Johnson JL, Tadesse DA, Hsu CH, Norris AL, Faust J, Walter-Grimm L, Colonius T: Artificial intelligence and machine learning in veterinary medicine: a regulatory perspective on current initiatives and future prospects . Am J Vet Res 2025, 86 (S1):S16-S21. Buono V, Sheikholharam Mashhadi P, Rahat M, Tiwari P, Byttner S: Expected Grad-CAM: Towards gradient faithfulness . arXiv e-prints 2024:arXiv:2406.01274. La Perle KMD: Machine Learning and Veterinary Pathology: Be Not Afraid! Vet Pathol 2019, 56 (4):506-507. Table Table 1. Dataset composition of canine cytological samples used in this study. Lymphoma RLH # Dog Training/Validation 85 81.0% 61 79.2% Test 20 19.0% 16 20.8% Total 105 77 # Image Training/Validation 1902 83.3% 2493 80.6% Test 382 16.7% 601 19.4% Total 2284 3094 Additional Declarations No competing interests reported. Supplementary Files SuppleTablesXAI.xlsx Cite Share Download PDF Status: Under Review Version 1 posted Reviewers agreed at journal 05 May, 2026 Reviewers agreed at journal 16 Apr, 2026 Reviewers invited by journal 01 Apr, 2026 Editor assigned by journal 26 Mar, 2026 Submission checks completed at journal 26 Mar, 2026 First submitted to journal 25 Mar, 2026 You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {"props":{"pageProps":{"initialData":{"identity":"rs-9225846","acceptedTermsAndConditions":true,"allowDirectSubmit":false,"archivedVersions":[],"articleType":"Research Article","associatedPublications":[],"authors":[{"id":634789896,"identity":"15b9d8c3-5923-43b8-bb85-605c3f6d46dd","order_by":0,"name":"Chanel Shum","email":"","orcid":"","institution":"University of Florida","correspondingAuthor":false,"prefix":"","firstName":"Chanel","middleName":"","lastName":"Shum","suffix":""},{"id":634789899,"identity":"12e36d42-f0bb-47c3-94d0-5c8b259dcea3","order_by":1,"name":"Donghee Lee","email":"","orcid":"","institution":"University of Florida","correspondingAuthor":false,"prefix":"","firstName":"Donghee","middleName":"","lastName":"Lee","suffix":""},{"id":634789900,"identity":"ce8c64f7-f85d-4f49-b289-5229ff8da9c5","order_by":2,"name":"Karthik Karnam","email":"","orcid":"","institution":"University of Florida","correspondingAuthor":false,"prefix":"","firstName":"Karthik","middleName":"","lastName":"Karnam","suffix":""},{"id":634789901,"identity":"f8b45285-7feb-4ec4-bdae-34e4b37a29e1","order_by":3,"name":"Matthew Vardas","email":"","orcid":"","institution":"University of Florida","correspondingAuthor":false,"prefix":"","firstName":"Matthew","middleName":"","lastName":"Vardas","suffix":""},{"id":634789904,"identity":"d3566baf-7946-463f-9a77-c1e756300c19","order_by":4,"name":"Samantha Masessa","email":"","orcid":"","institution":"University of Florida","correspondingAuthor":false,"prefix":"","firstName":"Samantha","middleName":"","lastName":"Masessa","suffix":""},{"id":634789905,"identity":"3739cb28-e6fa-4ec9-a052-96c785acaa40","order_by":5,"name":"Hyunji Jo","email":"","orcid":"","institution":"University of Florida","correspondingAuthor":false,"prefix":"","firstName":"Hyunji","middleName":"","lastName":"Jo","suffix":""},{"id":634789906,"identity":"fd135011-994b-4b00-bf07-67ee9969accf","order_by":6,"name":"Keerthana Chinthala","email":"","orcid":"","institution":"University of Florida","correspondingAuthor":false,"prefix":"","firstName":"Keerthana","middleName":"","lastName":"Chinthala","suffix":""},{"id":634789907,"identity":"a47e1b56-c5c7-4ffc-b9c6-ad6e1c0f29e5","order_by":7,"name":"Kevin Hall","email":"","orcid":"","institution":"University of Florida","correspondingAuthor":false,"prefix":"","firstName":"Kevin","middleName":"","lastName":"Hall","suffix":""},{"id":634789910,"identity":"f88041df-d80e-4eb9-bf17-2ede0c7ac837","order_by":8,"name":"Shir Gilor","email":"","orcid":"","institution":"University of Florida","correspondingAuthor":false,"prefix":"","firstName":"Shir","middleName":"","lastName":"Gilor","suffix":""},{"id":634789911,"identity":"14873cb0-bc6d-4e68-8b73-61e976c6a8e0","order_by":9,"name":"Jong Hyuk Kim","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAAArklEQVRIiWNgGAWjYBACxh7GBoaEChsG9nYgj4doLQ/OpDHwHCZWC0gV48O2wyRoYe453Lohse18Yg8zA+ODt23EOKy3se1GwrnbIC3MhnOJ0tLPCNRSdjtxPzMDmzQv8VrYzoFsYf9NnBaww9oOgLSwMROnpecgUMuZZOMeZsZmyTnniNBi2JP+7OaPCjvZHvbmgx/elBGjpQFhYQNOVShAnjhlo2AUjIJRMKIBAJi8OfIcI2xQAAAAAElFTkSuQmCC","orcid":"","institution":"University of Florida","correspondingAuthor":true,"prefix":"","firstName":"Jong","middleName":"Hyuk","lastName":"Kim","suffix":""},{"id":634789913,"identity":"3d5605d1-751d-4b1b-8341-0784ce192e39","order_by":10,"name":"Cleverson de Souza","email":"","orcid":"","institution":"University of Florida","correspondingAuthor":false,"prefix":"","firstName":"Cleverson","middleName":"","lastName":"de Souza","suffix":""}],"badges":[],"createdAt":"2026-03-25 17:09:27","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-9225846/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-9225846/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":108969388,"identity":"98c16b35-c310-4124-b238-c9f1e9d53d5b","added_by":"auto","created_at":"2026-05-11 10:10:30","extension":"png","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":51223853,"visible":true,"origin":"","legend":"\u003cp\u003ePhotomicrographs of canine lymphoma cytology. All panels show predominantly large neoplastic lymphocytes, with high N:C ratios, round nuclei approximately 2–3 RBCs in diameter, variable chromatin openness and up to three inconsistently conspicuous nucleoli. Background contains basophilic cytoplasmic fragments and occasional bare nuclei. \u003cstrong\u003eA:\u003c/strong\u003e Dog #1. The neoplastic lymphocytes are slightly more mixed in size, with mild amounts of deep blue cytoplasm, round nuclei with finely stippled to slightly coarse chromatin and 1-2 nucleoli. Low numbers of small, mature lymphocytes are present. \u003cstrong\u003eB:\u003c/strong\u003e Dog #2. Homogeneous neoplastic lymphocytes with mild amounts of deep blue cytoplasm, round nuclei with finely stippled chromatin and one large- or up to three small nucleoli. \u003cstrong\u003eC:\u003c/strong\u003e Dog #3. Homogeneous neoplastic lymphocytes with mild to moderate amounts of pale basophilic cytoplasm, round nuclei with fine chromatin and inconspicuous nucleoli. \u003cstrong\u003eD:\u003c/strong\u003e Dog #4. Homogeneous neoplastic lymphocytes with mild amounts of deep blue cytoplasm, round nuclei with fine chromatin and up to three nucleoli. Rare neutrophils and small mature lymphocytes are present. All images were taken at 1,000× magnification (10× ocular × 100× objective).\u003c/p\u003e","description":"","filename":"Fig1.png","url":"https://assets-eu.researchsquare.com/files/rs-9225846/v1/37c7ba7e3c23f51bfa8b0284.png"},{"id":108969399,"identity":"845262c4-8d1e-4573-9950-2b3eb95c1132","added_by":"auto","created_at":"2026-05-11 10:10:41","extension":"png","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":46254810,"visible":true,"origin":"","legend":"\u003cp\u003ePhotomicrographs of canine RLH. All panels show a heterogeneous lymphocyte population consisting of small, well differentiated lymphocytes, intermediate-size and occasional large immature lymphocytes. Background contains scattered cytoplasmic fragments. \u003cstrong\u003eA:\u003c/strong\u003e Dog #5. Mixed size lymphocytes consisting of small, intermediate, and fewer large lymphocytes. Small lymphocytes have a scant basophilic cytoplasm and mature, clumped chromatin with no nucleoli. Intermediate-size lymphocytes have slightly larger nuclei, mild amounts of basophilic cytoplasm, slightly coarse chromatin, and infrequently a faint nucleolus. Large lymphocytes have mild to moderate amounts of pale cytoplasm, large nuclei, with an open to lacey chromatin and 0 - 2 nucleoli. \u003cstrong\u003eB: \u003c/strong\u003eDog #6. Mixed small, intermediate, and fewer large lymphocytes. Overall, the cells have similar appearances to Fig 2a, except for few intermediate lymphocytes with a prominent, centrally located nucleolus, typical for marginal zone B lymphocytes. \u003cstrong\u003eC:\u003c/strong\u003e Dog #7. Small and intermediate, round and tapered appearing, lymphocytes predominate with scant to small amounts of basophilic cytoplasm and clumped, mature chromatin. There are rare plasma cells with a deeply basophilic cytoplasm, eccentric round nuclei and paranuclear clear zone. There is an eosinophil, bare nuclei and rare cytoplasmic fragments in the background. \u003cstrong\u003eD:\u003c/strong\u003eDog #8. Small and intermediate size well differentiated lymphocytes predominate. These have scant amounts of basophilic cytoplasm and coarsely clumped mature chromatin with no nucleoli. A large immature lymphocyte is present. It has a mild amount of deeply basophilic cytoplasm, round with a stippled chromatin and small inconspicuous nucleoli. All images were taken at 1,000× magnification (10× ocular × 100× objective).\u003c/p\u003e","description":"","filename":"Fig2.png","url":"https://assets-eu.researchsquare.com/files/rs-9225846/v1/4af7e081802d32e93c88e768.png"},{"id":108969387,"identity":"714d4862-49d0-45ca-8388-8c4d9c54635f","added_by":"auto","created_at":"2026-05-11 10:10:30","extension":"png","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":4153198,"visible":true,"origin":"","legend":"\u003cp\u003eTen‑fold cross‑validation training and validation performance of the classification model for canine lymphoma and RLH. \u003cstrong\u003eA, B:\u003c/strong\u003e Line graphs display the training (\u003cstrong\u003eA\u003c/strong\u003e) and validation (\u003cstrong\u003eB\u003c/strong\u003e) accuracy curves across 20 epochs for all ten folds. \u003cstrong\u003eC, D: \u003c/strong\u003eLine graphs show the corresponding training (\u003cstrong\u003eC\u003c/strong\u003e) and validation (\u003cstrong\u003eD\u003c/strong\u003e) loss line graphs across folds.\u003c/p\u003e","description":"","filename":"Fig3.png","url":"https://assets-eu.researchsquare.com/files/rs-9225846/v1/8290d2038671a093b3324d15.png"},{"id":108969697,"identity":"045e5d93-e34f-402a-8191-371ccb0ef77e","added_by":"auto","created_at":"2026-05-11 10:12:12","extension":"png","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":2079090,"visible":true,"origin":"","legend":"\u003cp\u003eTest performance of the trained model for binary classification of lymphoma versus RLH.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eA:\u003c/strong\u003e Table summarizes the final model performance on the hold-out test set. \u003cstrong\u003eB: \u003c/strong\u003eConfusion matrix illustrates the distribution of true and false positives and negatives. \u003cstrong\u003eC:\u003c/strong\u003eROC curve with the corresponding AUC value is shown.\u003c/p\u003e","description":"","filename":"Fig4.png","url":"https://assets-eu.researchsquare.com/files/rs-9225846/v1/6f96b26381e22a311c663591.png"},{"id":108969377,"identity":"f95a3c02-2c8b-41f8-8030-b9fbfda59804","added_by":"auto","created_at":"2026-05-11 10:10:25","extension":"png","order_by":5,"title":"Figure 5","display":"","copyAsset":false,"role":"figure","size":42123337,"visible":true,"origin":"","legend":"\u003cp\u003eVisual explanation of the model for classification of canine lymphoma. Representative original images (left panels) and corresponding Grad-CAM heatmaps (right panels) illustrate image regions contributing to the classification decisions. Paired images from four different canine patients demonstrate case-to-case variability while consistently highlighting similar regions, indicating intermediate to large lymphocytes identified as important by the model. Warmer colors (red/yellow) indicate regions of higher importance.\u003c/p\u003e","description":"","filename":"Fig5.png","url":"https://assets-eu.researchsquare.com/files/rs-9225846/v1/dd1e2b7895bdd4b9ed92a593.png"},{"id":108969376,"identity":"6521744f-0cb3-4bbe-9bf6-7dd1ea8b8a7f","added_by":"auto","created_at":"2026-05-11 10:10:25","extension":"png","order_by":6,"title":"Figure 6","display":"","copyAsset":false,"role":"figure","size":43321766,"visible":true,"origin":"","legend":"\u003cp\u003eVisual explanation of the model for classification of canine RLH. Representative original images (left panels) and corresponding Grad-CAM heatmaps (right panels) illustrate image regions contributing to the classification decisions. Paired images from four different canine patients demonstrate case-to-case variability while consistently highlighting similar regions, indicating small-sized lymphocytes or plasma cells identified as important by the model. Warmer colors (red/yellow) indicate regions of higher importance.\u003c/p\u003e","description":"","filename":"Fig6.png","url":"https://assets-eu.researchsquare.com/files/rs-9225846/v1/74fe676fcf9b231d1b9e0c5f.png"},{"id":108969719,"identity":"92702282-417e-4d76-9c59-ef83f852eba0","added_by":"auto","created_at":"2026-05-11 10:12:27","extension":"png","order_by":7,"title":"Figure 7","display":"","copyAsset":false,"role":"figure","size":24704941,"visible":true,"origin":"","legend":"\u003cp\u003eModel explanations for evaluating clinical applicability at the patient-level diagnosis. \u003cstrong\u003eA, B:\u003c/strong\u003e Bar plots show the proportions of correctly and incorrectly classified images among total test images per patient for lymphoma (A) and RLH (B). For lymphoma, true positives (TP) and false negatives (FN) are shown, while for RLH, true negatives (TN) and false positives (FP) are displayed. \u003cstrong\u003eC: \u003c/strong\u003eRepresentative photomicrographs show original images (left panels) and corresponding Grad-CAM heatmaps (right panels) from a lymphoma case (Patient #19) in which 19 of 23 images were misclassified. Red stars indicate neoplastic large lymphocytes, and yellow arrows indicate small and well differentiated lymphocytes. \u003cstrong\u003eD:\u003c/strong\u003eRepresentative photomicrographs show the original cytological morphology and corresponding Grad-CAM heatmaps for Patient #20, in which all four images were incorrectly classified. Black arrows indicate small mature lymphocytes.\u003c/p\u003e","description":"","filename":"Fig7.png","url":"https://assets-eu.researchsquare.com/files/rs-9225846/v1/e2ed6a85ee2ef6102624fc52.png"},{"id":108978290,"identity":"fb29f222-654a-46ae-9a4d-cb86db6dea80","added_by":"auto","created_at":"2026-05-11 11:35:58","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":179596761,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-9225846/v1/fed4e2e7-6627-4bfd-95bc-e785db90596f.pdf"},{"id":108969372,"identity":"ac6c1c65-e691-480c-a28e-45057422a927","added_by":"auto","created_at":"2026-05-11 10:10:25","extension":"xlsx","order_by":0,"title":"","display":"","copyAsset":false,"role":"supplement","size":18750,"visible":true,"origin":"","legend":"","description":"","filename":"SuppleTablesXAI.xlsx","url":"https://assets-eu.researchsquare.com/files/rs-9225846/v1/a577e883c517f7de1a04f55e.xlsx"}],"financialInterests":"No competing interests reported.","formattedTitle":"Applicability of explainable artificial intelligence in cytopathological evaluation of canine lymphoma","fulltext":[{"header":"Background","content":"\u003cp\u003eArtificial intelligence (AI) is a powerful tool for revolutionizing veterinary biomedical research, establishing a solid foundation for precision-driven digital health management in animal patients, especially companion animals [1, 2]. AI applications span disciplinary boundaries and may address gaps in veterinary knowledge [1, 3]. In particular, health management for companion animals often mirrors human medical systems, creating a bidirectional framework of reciprocal benefit. This convergence enables the development of clinically robust translational models that circumvent the ethical and biological limitations inherent in traditional laboratory animal research [4]. In this regard, the integration of AI and machine learning into veterinary oncology exemplifies the shift, offering scalable, data-intensive platforms that enhance diagnostic precision and accelerate therapeutic discovery. Yet, despite its transformative potential, the current stage of AI integration in veterinary medicine remains nascent. Successful clinical applications require rigorous validation and interdisciplinary collaboration to ensure reliability, interpretability, and clinical utility.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eLymphoma is the most prevalent hematopoietic malignancy in domestic dogs, comprising approximately 80% of all canine hematopoietic neoplasms\u0026nbsp;[5, 6]. Reactive lymphoid hyperplasia (RLH), a non-neoplastic lymph node enlargement characterized by a heterogeneous lymphoid population, represents the primary cytologic differential diagnosis for lymphoma\u0026nbsp;[7]. Initial diagnosis and treatment often begin in local veterinary clinics through physical examination and cytological evaluation of fine-needle aspirates. While histopathological examination offers the gold-standard for definite diagnosis\u0026nbsp;[5, 8], cytology offers a cost-effective and practical alternative, especially for cases where financial constraints or limited access to advanced diagnostics\u0026nbsp;pose challenges for pet owners and clinics\u0026nbsp;[8-11].\u0026nbsp;This approach is widely adopted in routine practice and often yields outcomes comparable to those achieved in tertiary veterinary centers. Fine-needle aspirate cytology is sufficient to establish a diagnosis in most dogs with typical multicentric lymphoma\u0026nbsp;[12]. However, excisional or core biopsy is recommended when cytological findings are equivocal, when low-grade lymphoma is suspected, or when full histological classification is required. Despite these guidelines, diagnostic and treatment decisions in local settings often rely on the anecdotal experience of general practitioners, particularly when input from clinical pathologists is unavailable. Standardized guidelines for cytological assessment\u0026nbsp;for lymphoma that\u0026nbsp;support more consistent evaluations are available to veterinary practitioners and medical oncologists\u0026nbsp;[7, 13]. These guidelines provide a framework for morphometric analysis of cytological images; however, they remain largely dependent on semi-quantitative assessments performed by the human visual system and require professional training. Recent advances and research endeavors\u0026nbsp;have demonstrated the application of deep learning techniques to canine lymphoma utilizing digital histopathological images\u0026nbsp;[14-17]. Nevertheless, clinically reliable digital systems for automated cytological analysis have yet to be established.\u003c/p\u003e\n\u003cp\u003eIn our previous study, we employed the ResNet-50 deep learning framework to evaluate the capability of pre-trained convolutional neural networks (CNNs) in classifying cytological images of canine lymphoma and RLH [18]. Despite the strong performance of the classification models, the study provided limited insights into model interpretability. Here, we present new approaches to refine the ResNet-50 models and to establish explainable AI frameworks that enhance interpretability and robustness in the cytopathological analysis of canine lymphoma.\u0026nbsp;\u003c/p\u003e"},{"header":"Methods","content":"\u003cp\u003e\u003cstrong\u003e\u003cem\u003eDatasets\u003c/em\u003e\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eWe utilized datasets of cytological images of canine lymphoma and RLH established in our previous study [18]. In the current study, these images were re-processed for data augmentation and mining to enhance their pathological relevance. Specifically, the 2,600 original images, prepared at 1920 \u0026times; 1080 pixels, were divided into four sub-images of 960 \u0026times; 540 pixels each, resulting in a total of 10,400 images. Color representation was standardized and images were normalized using ImageNet mean and standard deviation values of the RGB channels, as previously described [18]. Subsequently, all individual images underwent thorough review by three investigators: two clinical pathology residents (MV and CS) and one veterinary student (SM), all of whom were adequately trained in cytology image analysis by a board-certified clinical pathologist (CDS). During evaluation, we applied specific criteria to ensure the quality of images, focusing on those representing sufficient cellular components for pathological assessment and lymphoma diagnosis. The inclusion criteria required each image to contain more than 10 viable nucleated cells and more than seven viable lymphocytes. For images representing diagnosed lymphoma, only those containing at least 70% large lymphocytes among the total lymphocytes were included. Images of poor quality, such as those that were out of focus or had excessively high cell density preventing individual cell distinction, were excluded. This quality control review produced image datasets with adequate components per image, suitable for training purposes to differentiate between lymphoma and RLH. The final image dataset consisted of 5,378 images, including 2,284 images of lymphoma (105 dogs) and 3,094 images of RLH (77 dogs). Representative images for each group are provided (\u003cstrong\u003eFigs. 1 and 2\u003c/strong\u003e).\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e\u003cem\u003eProgramming environment\u003c/em\u003e\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eAll experiments were conducted in a Python (version 3.10.12) environment on a Red Hat Enterprise Linux (RHEL) 8 operating system. The deep learning models were implemented, trained, and evaluated using PyTorch (version 2.5.1) and its accompanying torchvision library (version 0.20.1) for model access and image transformations. GPU acceleration was provided by NVIDIA A100 GPUs (80 GB VRAM each), utilizing the CUDA Toolkit (version 11.8). Key supporting libraries included scikit-learn (version 1.4.0) for calculating performance metrics, NumPy (version 1.26.4) for numerical operations, and Pillow (version 11.0.0) for image loading and initial manipulation. All computational tasks were performed on the University of Florida\u0026apos;s research computing cluster, HiPerGator, equipped with an AMD EPYC 75F3 Milan CPU (64 cores) and 500 GB of RAM.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e\u003cem\u003eModel development\u003c/em\u003e\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eData splitting and model training were performed as previously described [18]. Briefly, digital images that passed quality control were divided at the dog-patient level into training and hold-out test sets in an 8:2 ratio. Partitioning was done by randomly shuffling patients using numpy.random.permutation (NumPy version 1.26.4) with a fixed random seed. Approximately 20% of patients from each disease group were allocated to the test set, with the remaining 80% used for training and validation. \u003cstrong\u003eTable 1\u003c/strong\u003e summarizes the dataset composition, including the numbers of dog patients and images. Additional data augmentation including random vertical and horizontal flipping was conducted using torchvision.transforms.RandomVerticalFlip(p=0.5) and torchvision.transforms.RandomHorizontalFlip(p=0.5) (Torchvision version 0.20.1). Random 90\u0026deg; rotation and color jitter (brightness \u0026plusmn; 20%, contrast \u0026plusmn; 20%, saturation \u0026plusmn; 10%, hue \u0026plusmn; 0.02) were also applied.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eTransfer learning was then performed on the training/validation datasets using the ResNet-50 CNN architecture initialized with weights pretrained on ImageNet-1K (ILSVRC2012) [19]. The final fully connected layer was replaced with a single-unit linear layer to enable binary classification, followed by sigmoid activation to generate probability outputs. Hyperparameter tuning was conducted via 10-fold cross-validation. For this process, the datasets excluding the test set were randomly divided at the dog-patient level into ten distinct groups in a 9:1 ratio for training and validation subsets. Training and validation were repeated across all ten folds. Early stopping was employed by monitoring validation loss, with training halted if no improvement was observed for four consecutive epochs (patience = 4), up to a maximum of 50 epochs. The model was trained using the Adam optimizer (learning rate = 1e-4, no scheduler, no weight decay) with a batch size of 2. Binary cross-entropy loss was used without class weighting. All layers were kept unfrozen during training. After each fold, accuracy and loss were computed and recorded for both training and validation sets.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e\u003cem\u003eModel evaluation\u003c/em\u003e\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eAfter hyperparameter optimization through 10-fold cross-validation on the training/validation datasets, we evaluated the model on the hold-out test set (382 lymphoma images from 20 dogs; 601 RLH images from 16 dogs). Performance was assessed at two levels: image- and patient-level. The performance metrics included accuracy, weighted average precision, weighted average recall, weighted average F1-score, and the area under the receiver operating characteristic curve (AUC-ROC), and the area under the precision-recall curve (AUC-PR). Considering the imbalanced distribution of the datasets among the classification groups, model performance for each fold was evaluated using this comprehensive set of metrics, alongside training and validation curves, as previously described [20, 21]. 95% confidence intervals (CIs) were estimated using the Wilson score method. Image-level predictions were generated using a classification threshold of 0.5, and a majority-vote rule was applied to determine the classification of each dog at the patient level based on the predominant image-level predictions.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e\u003cem\u003eGradient-weighted Class Activation Mapping (Grad-CAM)\u003c/em\u003e\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eGrad-CAM was employed to provide visual explanations from our developed model [22]. Visualizations were generated using the \u003cem\u003epytorch-grad-cam\u003c/em\u003e implementation applied to the last convolutional block of ResNet-50. For each image, class-specific gradients were back-propagated to this layer, gradients were globally averaged to weight feature maps, the weighted sum was passed through ReLU, normalized, upsampled to the input resolution, and overlaid on the original RGB image. Grad-CAMs were produced during evaluation for representative cases and exported in batch for the entire test set, and outputs were saved to the results directory while preserving class and patient subfolders with a manifest CSV of file paths and labels.\u003c/p\u003e"},{"header":"Results","content":"\u003cp\u003e\u003cstrong\u003e\u003cem\u003eResNet-50-based models show robust and consistent learning for binary classification of canine lymphoma and RLH cytological images\u003c/em\u003e\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eFirst, we developed ten deep learning models using 1,902 lymphoma images from 85 dogs and 2,493 RLH images from 61 dogs through 10-fold cross-validation. We found that the models exhibited robust and consistent learning behavior across folds. On average, training accuracy increased from 0.826 to 0.972, while validation accuracy improved from 0.887 to 0.975 (\u003cstrong\u003eFigs. 3A, B\u003c/strong\u003e). By the final optimized epoch, training accuracy remained consistently high across all folds, reaching between 0.951 and 0.970, indicating strong model fit to the training data. Validation accuracy improved steadily with each epoch, with most folds achieving final values between 0.907 and 0.986, reflecting robust generalization. Training loss decreased from 0.390 to 0.074, and validation loss followed a comparable downward trend, albeit with slightly higher values and greater variability than training loss (\u003cstrong\u003eFigs. 3C, D\u003c/strong\u003e). Training loss declined consistently across epochs in all folds, often reaching values below 0.10. Validation loss also decreased in most folds, though some exhibited plateaus or fluctuations. Collectively, the training and validation curves across all folds suggest effective learning, convergence, and minimal overfitting. Complete metrics of training and validation performance over epochs using 10-fold cross-validation are provided in \u003cstrong\u003eSupplementary Table S1\u003c/strong\u003e.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e\u003cem\u003eEvaluation of test performance demonstrates model suitability and clinically relevant insights into diagnostic utility.\u003c/em\u003e\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eWith optimized hyperparameters, we assessed a final model performance using a hold-out test dataset comprising 382 lymphoma images from 20 dogs and 601 RLH images from 16 dogs. The performance metrics included accuracy, weighted precision, weighted recall, weighted F1-score, AUC-ROC, and AUC-PR. First, at the image level, with patient-level dataset splitting, the model demonstrated a test accuracy of 0.945 (95% CI: 0.929 - 0.958), a weighted precision of 0.945 (95% CI: 0.903 - 0.939), a weighted recall of 0.945 (95% CI: 0.899 - 0.969), a weighted F1-score of 0.945 (95% CI: 0.910 - 0.943), an AUC-ROC of 0.979 (95% CI: 0.957 - 0.978), and an AUC-PR of 0.981 (95% CI: 0.960 - 0.985) (\u003cstrong\u003eFig. 4\u003c/strong\u003e). Then, to assess clinically meaningful diagnostic performance, we evaluated the model at the patient level using a majority vote rule, classifying each dog based on the predominant image-level predictions. Under this rule, 18 of 20 lymphoma dogs were correctly classified (true positives), and all 16 RLH dogs were correctly classified (true negatives). This corresponded to a sensitivity of 0.900 (95% CI: 0.699 - 0.972), specificity of 1.000 (95% CI: 0.806 - 1.000), positive predictive value of 1.000 (95% CI: 0.824 - 1.000), and negative predictive value of 0.889 (95% CI: 0.672 - 0.969).\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eThese findings indicate that our ResNet-50-based model is well-suited for binary classification of canine lymphoma and RLH cytological images. Also, patient-level performance evaluation provides clinically relevant insights into diagnostic utility.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e\u003cem\u003eGrad-CAM supports morphological identification of malignant lymphocytes distinct from benign cells\u0026nbsp;\u003c/em\u003e\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eTo address model explainability, we deployed Grad-CAM, which visualizes the decision-making features within the cytological images. We generated Grad-CAM heatmaps for all test set images from both groups. The heatmaps highlighted regions enriched with large lymphocytes in lymphoma images (\u003cstrong\u003eFig. 5\u003c/strong\u003e) and regions enriched with small mature lymphocytes and rare plasma cells in RLH images (\u003cstrong\u003eFig. 6\u003c/strong\u003e). Despite considerable morphological diversity, cellular composition, and color variation across groups, these Grad-CAM findings demonstrate that the model extracted deterministic features differentiating lymphoma from RLH, concordant with cellular characteristics used for diagnostic identification.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eAs described above, test performance was evaluated at both the image and patient level. To further enhance the model interpretability, we calculated the proportions of correctly classified images for each dog patient. Among lymphoma cases, 18 of 20 dogs had more than 66.7% of their images correctly classified; specifically, 16 dogs exceeded 97% and two dogs had 84.6% and 66.7%, respectively (\u003cstrong\u003eFig. 7A; Supplementary Table S2\u003c/strong\u003e). Two lymphoma cases (Patients #19 and #20) demonstrated markedly lower correct classification rates of 17.4% and 0%. For RLH cases, all 16 dogs had at least 70% of their images correctly identified, representing true negatives (\u003cstrong\u003eFig. 7B; Supplementary Table S3\u003c/strong\u003e). To investigate the two lymphoma cases with high rates of misclassified images (i.e., false negatives), we examined the Grad-CAM results.\u0026nbsp;In Patient #19, cytological evaluation revealed an abnormal dimorphic lymphocyte population characterized by small mature lymphocytes and an expanded population of large immature lymphocytes, with an absence of intermediate-size lymphocytes, findings that are morphologically diagnostic of lymphoma (\u003cstrong\u003eFig. 7C\u003c/strong\u003e). However, only 4 of 23 images were correctly predicted, with the Grad-CAM highlighting regions containing large lymphocytes. The other 19 images were misclassified as RLH, with the Grad-CAM highlighting regions containing small or normal-sized lymphocytes. This pattern suggests that our model distinguishes lymphoma and RLH based on differences in cell size between large and small lymphocytes. In Patient #20, all four images were misclassified despite meeting all pre-processing criteria; they showed relatively low cellularity (\u003cstrong\u003eFig. 7D\u003c/strong\u003e). Grad-CAM highlighted areas containing small lymphocytes, suggesting that samples with reduced cellularity and heterogeneous morphology across different areas may also contribute to false-negative predictions.\u003c/p\u003e\n\u003cp\u003eCollectively, our observations suggest that Grad-CAM technique may aid model interpretability by highlighting image regions, although these visualizations should be interpreted as coarse regional associations rather than cell-specific explanations.\u003c/p\u003e"},{"header":"Discussion","content":"\u003cp\u003eThis study shows that a ResNet-50 deep learning architecture facilitates reliable model development for AI-assisted identification of canine lymphoma using cytological images. It also provides a useful approach for integrating explainable AI tools to enhance model explainability in the context of cytological characterization. Our work presents a computational workflow that mirrors key aspects of the morphological differentiation of malignant lymphoid cells from benign conditions in canine specimens, thereby supporting technological advancement in veterinary cytopathology.\u003c/p\u003e\n\u003cp\u003eResNet-50 is a widely utilized deep learning architecture in computer vision [23]. Its effectiveness, along with strong support from both the scientific community and industry, has made pretrained ResNet-50 models a popular choice for transfer learning [24-27]. This widespread adoption offers practical advantages to developers and users, such as robust technical resources and guidance for mitigating bias during model development and subsequent refinement [28]. Although AI-driven image classification has become popular, deep learning applications in veterinary cytology are still limited, at least in part, due to restricted technical accessibility in the veterinary field. Tasks require advanced computational expertise and high-performance hardware. However, these barriers are rapidly receding with ongoing advances in computing power and development tools. Cytological examinations rely on holistic morphological assessment by expert clinical pathologists. Consequently, employing ResNet-50 for cytological image analysis represents a viable strategy, with reasonable prospects for success, provided the images contain sufficient morphological features for the model to learn. While this potential has been explored in human clinical cytopathology [29-31], we recently extended it to veterinary medicine. We previously developed prototype models to differentiate canine lymphoma from RLH, and to distinguish B-cell from T-cell lymphoma using a dataset of 2,600 images captured at 1920 \u0026times; 1080 pixels [18]. Despite rigorous technical validation and comprehensive performance evaluation, this prior work lacked model explainability, possibly because cell-level segmentation and annotation were not performed. In the current study, we continued to employ weakly supervised, image-level annotation, but supplemented it with rigorous human intervention to enhance model interpretability. Specifically, a dataset of 5,378 images, pre-processed at 960 \u0026times; 540 pixels, underwent additional quality control by clinical pathologists to ensure that each image contained a sufficient number of viable nucleated cells and large lymphocytes.\u003c/p\u003e\n\u003cp\u003eThe assessment of individual images was not definitive for lymphoma diagnosis; in practice, pathologists review entire slides in conjunction with clinical history. Accordingly, our deep learning models do not yet fully replicate the comprehensive diagnostic process performed by clinicians. The goal of this study was not to replace pathologists but to evaluate whether ResNet-50 deep learning architecture has practical utility for cytological image analysis in veterinary oncology [32, 33]. Our results suggest that transfer learning with pretrained ResNet-50 models is feasible for canine lymphoma cytology, supporting both the versatility and broader applicability of this approach. By emphasizing transparency in model development, this work addresses concerns related to unclear or opaque procedures that can undermine reliability and trust in AI models in the veterinary domain [33-36]. The integration of explainable AI tools further strengthens this effort: Grad-CAM visualizations provided evidence of differential morphological features captured by the models between lymphoma and RLH cases. Intriguingly, similar to human visual assessment, the models appeared to focus on image regions that reflect features typically used to differentiate benign from malignant conditions. However, reliance on a single explainability method such as Grad-CAM may raise concerns regarding faithfulness and completeness, and alternative explainable AI techniques are available [37]. Thus, future studies should include comparative evaluations of multiple explainability approaches across different model architectures.\u003c/p\u003e\n\u003cp\u003eThe misclassified lymphoma cases are unlikely random errors but reflect systematic vulnerabilities of the model. Plasma cell-rich or plasmacytoid variants represent a well-recognized diagnostic challenge, even for experienced pathologists [9]. Similarly, low-cellularity samples, which can result from hemodilution or suboptimal aspiration technique, are a common pre-analytical issue in cytology and may be more frequent among general practitioners. At the patient level, the model exhibited a 10% false-negative rate (2/20 lymphoma cases), with a negative predictive value of 0.889, indicating that approximately 1 in 9 cases predicted as \u0026ldquo;negative\u0026rdquo; may in fact represent lymphoma. Clinically, such false negatives could delay diagnosis and treatment, underscoring the need for cautious interpretation of negative results, particularly in cases with atypical morphology or low cellularity. Potential mitigation strategies include implementing a minimum confidence threshold for automated predictions or incorporating a pre-screening cellularity check to flag low-cellularity images for manual review, thereby improving safety in real-world deployment.\u003c/p\u003e\n\u003cp\u003eA key limitation of this study is that the model classifies only lymphoma versus RLH. In clinical practice, lymph node enlargement may result from a broad range of conditions, including poorly cohesive metastatic carcinoma, melanoma, mast cell tumor, histiocytic sarcoma, granulomatous or suppurative lymphadenitis, and others. As a closed-set binary classifier, the model necessarily assigns every image to one of the two predefined categories. A lymph node containing metastatic neoplasia that is misclassified as RLH would represent a clinically significant false negative that the current evaluation framework cannot detect. Future work should explore multi-class classification models or open-set recognition approaches to better reflect real-world diagnostic complexity.\u003c/p\u003e\n\u003cp\u003eBeyond technical and clinical performance, it is essential to address the professional skepticism and heightened caution prevalent within the veterinary community [33, 34, 38]. The primary barrier to AI adoption is often not the technology itself, but a lack of transparency and rigor during development. From a pathologist\u0026rsquo;s perspective, AI still requires significant human oversight, including case review, quality control, data collection, in-depth cellular segmentation, and micrometer-resolution annotation. Despite the necessary involvement of human experts, scientific publications and commercial products often describe AI capabilities in ways that can seem overstated or unclear. This disparity creates a lack of accountability that hinders clinical trust. By prioritizing rigorous technical validation and model explainability, our study provides the transparency necessary to mitigate these concerns. Although the FDA and other bodies have yet to establish comprehensive regulatory frameworks to keep pace with these innovations [36], technical transparency and regulatory guidance must advance in tandem to ensure the safe and reliable integration of AI into veterinary medicine.\u003c/p\u003e\n\u003cp\u003eIn conclusion, our study demonstrates a feasible and transparent approach for developing AI-assisted cytological classification in veterinary medicine. By moving beyond \u0026ldquo;black-box\u0026rdquo; methodologies and incorporating human-verified datasets and Grad-CAM visualizations, our fundings suggest that AI systems can be developed with rigor appropriate for preliminary clinical evaluation. To further refine these models and assess the clinical reliability, multi-institutional studies with large research consortia are needed. While these models do not replace the holistic expertise of a pathologist, they offer a powerful digitized platform for enhancing diagnostic workflows. As technical and regulatory guidance continues to advance in tandem, this transparent developmental model provides a template for the responsible and trustworthy implementation of AI in veterinary practice.\u003c/p\u003e"},{"header":"Abbreviations","content":"\u003cp\u003eAI: artificial intelligence\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eAUC-PR: area under the precision-recall curve\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eAUC-ROC: area under the receiver operating characteristic curve\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eCI: confidence interval\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eCNN: convolutional neural networks\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eCPU: central processing unit\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eCSV: comma-separated values\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eCUDA: compute unified device architecture\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eGPU: graphics processing unit\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eGrad-CAM: gradient-weighted class activation mapping\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eIACUC: institutional animal care and use committee\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eRAM: random access memory\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eRGB: red-green-blue\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eRHEL: Red Hat Enterprise Linux\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eRLH: reactive lymphoid hyperplasia\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eVRAM: video random access memory\u0026nbsp;\u003c/p\u003e"},{"header":"Declarations","content":"\u003cp\u003e\u003cstrong\u003eEthics approval and consent to participate\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThis retrospective study analyzed existing, de-identified cytology images from client-owned dogs. Veterinary clinical samples were collected under standard clinical care with owner consent in accordance with institutional policy, or institutional review determined that consent was not required for de-identified retrospective data. No prospective enrollment, procedures, or sample collection from live animals were performed; thus, formal Institutional Animal Care and Use Committee (IACUC) review was not required. This determination is consistent with PHS/OLAW guidance on the scope of IACUC oversight.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eConsent for publication\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eNot applicable.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eAvailability of data and materials\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe source code used for image preprocessing, model training, and evaluation in this study will be made publicly available on GitHub at [Placeholder for URL \u0026ndash; e.g., to be provided upon acceptance and code upload]. The anonymized cytological image dataset supporting the findings of this study is available from the corresponding author (J.H.K and C.S.) upon reasonable request, subject to institutional guidelines and a data use agreement.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eCompeting interests\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe authors declare that they have no competing interests.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eFunding\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe authors disclosed the following financial support for the research, authorship, and publication of this article: this work was partially supported by the CVM Clinical Pathology Residency program, the UF Strategic Funding Initiative, AI Initiative, AI Scholar Award, and Lisa Conti One Health Initiative through the Florida Veterinary Scholars Program (FVSP) at the University of Florida. This work was also supported by the Mid-Florida Golden Retriever Club.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eAuthors\u0026apos; contributions\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eJHK and CDS designed the study and protocols; CS, MV, SM, KH, SG, and CDS collected the samples, performed microscopic evaluations, and captured images; DL, KK, HJ, and KC performed data preprocessing and computational work; KH contributed to administrative support; the manuscript was written by CS, DL, KK, MV, JHK, and CDS with contribution from all other authors.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eAcknowledgements\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eWe would like to acknowledge the staff and technicians at the clinical pathology service for supporting sample collection and data curation at the University of Florida Small Animal Hospital. The authors also acknowledge computational resources and technical support provided by the HiPerGator Research Computing at the University of Florida.\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\n\u003cli\u003eAkinsulie OC, Idris I, Aliyu VA, Shahzad S, Banwo OG, Ogunleye SC, Olorunshola M, Okedoyin DO, Ugwu C, Oladapo IP\u003cem\u003e et al\u003c/em\u003e: \u003cstrong\u003eThe potential application of artificial intelligence in veterinary clinical practice and biomedical research\u003c/strong\u003e. \u003cem\u003eFront Vet Sci \u003c/em\u003e2024, \u003cstrong\u003e11\u003c/strong\u003e:1347550.\u003c/li\u003e\n\u003cli\u003eArshad MF, Ahmed F, Nonnis F, Tamponi C, Scala A, Varcasia A: \u003cstrong\u003eArtificial intelligence and companion animals: Perspectives on digital healthcare for dogs, cats, and pet ownership\u003c/strong\u003e. \u003cem\u003eRes Vet Sci \u003c/em\u003e2025, \u003cstrong\u003e193\u003c/strong\u003e:105776.\u003c/li\u003e\n\u003cli\u003eBasran PS, Appleby RB: \u003cstrong\u003eThe unmet potential of artificial intelligence in veterinary medicine\u003c/strong\u003e. \u003cem\u003eAm J Vet Res \u003c/em\u003e2022, \u003cstrong\u003e83\u003c/strong\u003e(5):385-392.\u003c/li\u003e\n\u003cli\u003eOh JH, Cho JY: \u003cstrong\u003eComparative oncology: overcoming human cancer through companion animal studies\u003c/strong\u003e. \u003cem\u003eExp Mol Med \u003c/em\u003e2023, \u003cstrong\u003e55\u003c/strong\u003e(4):725-734.\u003c/li\u003e\n\u003cli\u003eZandvliet M: \u003cstrong\u003eCanine lymphoma: a review\u003c/strong\u003e. \u003cem\u003eVet Q \u003c/em\u003e2016, \u003cstrong\u003e36\u003c/strong\u003e(2):76-104.\u003c/li\u003e\n\u003cli\u003eVail D, Thamm D, Liptak J: \u003cstrong\u003eHematopoietic Tumors.\u003c/strong\u003e In: \u003cem\u003eWithrow and MacEwen\u0026apos;s Small Animal Clinical Oncology.\u003c/em\u003e edn.; 2019: 688-772.\u003c/li\u003e\n\u003cli\u003eRaskin RE, Meyer D: \u003cstrong\u003eCanine and Feline Cytology\u0026mdash;A Color Atlas and Interpretation Guide\u003c/strong\u003e, 2nd edn. St. Louis, MO: Elsevier; 2009.\u003c/li\u003e\n\u003cli\u003eSeelig DM, Avery AC, Ehrhart EJ, Linden MA: \u003cstrong\u003eThe Comparative Diagnostic Features of Canine and Human Lymphoma\u003c/strong\u003e. \u003cem\u003eVet Sci \u003c/em\u003e2016, \u003cstrong\u003e3\u003c/strong\u003e(2).\u003c/li\u003e\n\u003cli\u003eAresu L: \u003cstrong\u003eCanine Lymphoma, More Than a Morphological Diagnosis: What We Have Learned about Diffuse Large B-Cell Lymphoma\u003c/strong\u003e. \u003cem\u003eFront Vet Sci \u003c/em\u003e2016, \u003cstrong\u003e3\u003c/strong\u003e:77.\u003c/li\u003e\n\u003cli\u003eRiondato F, Comazzi S: \u003cstrong\u003eFlow Cytometry in the Diagnosis of Canine B-Cell Lymphoma\u003c/strong\u003e. \u003cem\u003eFront Vet Sci \u003c/em\u003e2021, \u003cstrong\u003e8\u003c/strong\u003e:600986.\u003c/li\u003e\n\u003cli\u003eDas DK: \u003cstrong\u003eContribution of Immunocytochemistry to the Diagnosis of Usual and Unusual Lymphoma Cases\u003c/strong\u003e. \u003cem\u003eJ Cytol \u003c/em\u003e2018, \u003cstrong\u003e35\u003c/strong\u003e(3):163-169.\u003c/li\u003e\n\u003cli\u003eChildress MO, Avery A, Behling-Kelly E, Bennett P, Brockley L, Dickinson R, Hughes K, Kisseberth WC, Marconato L, Martini V\u003cem\u003e et al\u003c/em\u003e: \u003cstrong\u003eDiagnosis and Classification of Primary Nodal Lymphomas in Dogs: A Consensus of the Oncology-Pathology Working Group\u003c/strong\u003e. \u003cem\u003eVet Comp Oncol \u003c/em\u003e2025, \u003cstrong\u003e23\u003c/strong\u003e(3):331-345.\u003c/li\u003e\n\u003cli\u003eValenciano AC, Cowell RL: \u003cstrong\u003eCowell and Tyler\u0026rsquo;s Diagnostic Cytology and Hematology of the Dog and Cat\u003c/strong\u003e, 5th edn. St. Louis, MO: Elsevier; 2020.\u003c/li\u003e\n\u003cli\u003eHubbard-Perez M, Luchian A, Milford C, Ressel L: \u003cstrong\u003eUse of deep learning for the classification of hyperplastic lymph node and common subtypes of canine lymphomas: a preliminary study\u003c/strong\u003e. \u003cem\u003eFront Vet Sci \u003c/em\u003e2023, \u003cstrong\u003e10\u003c/strong\u003e:1309877.\u003c/li\u003e\n\u003cli\u003eHaghofer A, Fuchs-Baumgartinger A, Lipnik K, Klopfleisch R, Aubreville M, Scharinger J, Weissenb\u0026ouml;ck H, Winkler SM, Bertram CA: \u003cstrong\u003eHistological classification of canine and feline lymphoma using a modular approach based on deep learning and advanced image processing\u003c/strong\u003e. \u003cem\u003eSci Rep \u003c/em\u003e2023, \u003cstrong\u003e13\u003c/strong\u003e(1):19436.\u003c/li\u003e\n\u003cli\u003eAncheta K, Le Calvez S, Williams J: \u003cstrong\u003eThe digital revolution in veterinary pathology\u003c/strong\u003e. \u003cem\u003eJ Comp Pathol \u003c/em\u003e2024, \u003cstrong\u003e214\u003c/strong\u003e:19-31.\u003c/li\u003e\n\u003cli\u003eAncheta K, Psifidi A, Yale AD, Le Calvez S, Williams J: \u003cstrong\u003eDeep-learning based morphological segmentation of canine diffuse large B-cell lymphoma\u003c/strong\u003e. \u003cem\u003eFront Vet Sci \u003c/em\u003e2025, \u003cstrong\u003e12\u003c/strong\u003e:1656976.\u003c/li\u003e\n\u003cli\u003eShum C, Lee D, Jo H, Kim J, Heo H, Gilor S, Lanier C, Hall K, Dark M, Kim JH\u003cem\u003e et al\u003c/em\u003e: \u003cstrong\u003eDeep learning-assisted cytological image analysis for canine lymphoma\u003c/strong\u003e. \u003cem\u003eVeterinary Oncology \u003c/em\u003e2026, \u003cstrong\u003e3\u003c/strong\u003e(1):5.\u003c/li\u003e\n\u003cli\u003eRussakovsky O, Deng J, Su H, Krause J, Satheesh S, Ma S, Huang Z, Karpathy A, Khosla A, Bernstein M\u003cem\u003e et al\u003c/em\u003e: \u003cstrong\u003eImageNet Large Scale Visual Recognition Challenge.\u003c/strong\u003e \u003cem\u003eInt J Comput Vis \u003c/em\u003e2015, \u003cstrong\u003e115\u003c/strong\u003e(3):211\u0026ndash;252.\u003c/li\u003e\n\u003cli\u003eMaier-Hein L, Reinke A, Godau P, Tizabi MD, Buettner F, Christodoulou E, Glocker B, Isensee F, Kleesiek J, Kozubek M\u003cem\u003e et al\u003c/em\u003e: \u003cstrong\u003eMetrics reloaded: recommendations for image analysis validation\u003c/strong\u003e. \u003cem\u003eNat Methods \u003c/em\u003e2024, \u003cstrong\u003e21\u003c/strong\u003e(2):195-212.\u003c/li\u003e\n\u003cli\u003eGaudreault J-G, Branco P: \u003cstrong\u003eEmpirical analysis of performance assessment for imbalanced classification\u003c/strong\u003e. \u003cem\u003eMachine Learning \u003c/em\u003e2024, \u003cstrong\u003e113\u003c/strong\u003e(8):5533-5575.\u003c/li\u003e\n\u003cli\u003eSelvaraju RR, Cogswell M, Das A, Vedantam R, Parikh D, Batra D: \u003cstrong\u003eGrad-CAM: Visual Explanations from Deep Networks via Gradient-Based Localization\u003c/strong\u003e. In: \u003cem\u003e2017 IEEE International Conference on Computer Vision (ICCV).\u003c/em\u003e Venice, Italy; 2017: 618-626.\u003c/li\u003e\n\u003cli\u003eHe K, Zhang X, Ren S, Sun J: \u003cstrong\u003eDeep Residual Learning for Image Recognition\u003c/strong\u003e. In: \u003cem\u003e2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR): 27-30 June 2016 2016\u003c/em\u003e; 2016: 770-778.\u003c/li\u003e\n\u003cli\u003eZaidi SSA, Ansari MS, Aslam A, Kanwal N, Asghar M, Lee B: \u003cstrong\u003eA survey of modern deep learning based object detection models\u003c/strong\u003e. \u003cem\u003eDigital Signal Processing \u003c/em\u003e2022, \u003cstrong\u003e126\u003c/strong\u003e:103514.\u003c/li\u003e\n\u003cli\u003eDesanamukula VS, Teja TD, Rajitha P: \u003cstrong\u003eAn In-Depth Exploration of ResNet-50 and Transfer Learning in Plant Disease Diagnosis\u003c/strong\u003e. In: \u003cem\u003e2024 International Conference on Inventive Computation Technologies (ICICT): 24-26 April 2024 2024\u003c/em\u003e; 2024: 614-621.\u003c/li\u003e\n\u003cli\u003eAlzubaidi L, Zhang J, Humaidi AJ, Al-Dujaili A, Duan Y, Al-Shamma O, Santamar\u0026iacute;a J, Fadhel MA, Al-Amidie M, Farhan L: \u003cstrong\u003eReview of deep learning: concepts, CNN architectures, challenges, applications, future directions\u003c/strong\u003e. \u003cem\u003eJ Big Data \u003c/em\u003e2021, \u003cstrong\u003e8\u003c/strong\u003e(1):53.\u003c/li\u003e\n\u003cli\u003eJena B, Nayak GK, Saxena S: \u003cstrong\u003eConvolutional neural network and its pretrained models for image classification and object detection: A survey\u003c/strong\u003e. \u003cem\u003eConcurr Comput Pract Exp \u003c/em\u003e2022, \u003cstrong\u003e34\u003c/strong\u003e(6):e6767.\u003c/li\u003e\n\u003cli\u003eHuang Y, Lin L, Cheng P, Lyu J, Tam R, Tang X: \u003cstrong\u003eIdentifying the Key Components in ResNet-50 for Diabetic Retinopathy Grading from Fundus Images: A Systematic Investigation\u003c/strong\u003e. \u003cem\u003eDiagnostics (Basel) \u003c/em\u003e2023, \u003cstrong\u003e13\u003c/strong\u003e(10).\u003c/li\u003e\n\u003cli\u003eSu F, Sun Y, Hu Y, Yuan P, Wang X, Wang Q, Li J, Ji J-F: \u003cstrong\u003eDevelopment and validation of a deep learning system for ascites cytopathology interpretation\u003c/strong\u003e. \u003cem\u003eGastric Cancer \u003c/em\u003e2020, \u003cstrong\u003e23\u003c/strong\u003e(6):1041-1050.\u003c/li\u003e\n\u003cli\u003eKaur H, Sharma R, Kaur J: \u003cstrong\u003eComparison of deep transfer learning models for classification of cervical cancer from pap smear images\u003c/strong\u003e. \u003cem\u003eSci Rep \u003c/em\u003e2025, \u003cstrong\u003e15\u003c/strong\u003e(1):3945.\u003c/li\u003e\n\u003cli\u003eRodr\u0026iacute;guez M, C\u0026oacute;rdova C, Benjumeda I, San Mart\u0026iacute;n S: \u003cstrong\u003eAutomated Cervical Cancer Screening Using Single-Cell Segmentation and Deep Learning: Enhanced Performance with Liquid-Based Cytology\u003c/strong\u003e. \u003cem\u003eComputation \u003c/em\u003e2024, \u003cstrong\u003e12\u003c/strong\u003e(12):232.\u003c/li\u003e\n\u003cli\u003ePacholec C, Flatland B, Xie H, Zimmerman K: \u003cstrong\u003eHarnessing artificial intelligence for enhanced veterinary diagnostics: A look to quality assurance, Part II External validation\u003c/strong\u003e. \u003cem\u003eVet Clin Pathol \u003c/em\u003e2025.\u003c/li\u003e\n\u003cli\u003eSobkowich KE: \u003cstrong\u003eDemystifying artificial intelligence for veterinary professionals: practical applications and future potential\u003c/strong\u003e. \u003cem\u003eAm J Vet Res \u003c/em\u003e2025, \u003cstrong\u003e86\u003c/strong\u003e(S1):S6-S15.\u003c/li\u003e\n\u003cli\u003eCazer CL, Basran P, Ivanek-Miojevic R: \u003cstrong\u003eFrom bark to bytes: artificial intelligence transforming veterinary medicine\u003c/strong\u003e. \u003cem\u003eAm J Vet Res \u003c/em\u003e2025, \u003cstrong\u003e86\u003c/strong\u003e(S1):S4-S5.\u003c/li\u003e\n\u003cli\u003eCohen EB, Gordon IK: \u003cstrong\u003eFirst, do no harm. Ethical and legal issues of artificial intelligence and machine learning in veterinary radiology and radiation oncology\u003c/strong\u003e. \u003cem\u003eVet Radiol Ultrasound \u003c/em\u003e2022, \u003cstrong\u003e63 Suppl 1\u003c/strong\u003e(Suppl 1):840-850.\u003c/li\u003e\n\u003cli\u003eDuggirala HJ, Johnson JL, Tadesse DA, Hsu CH, Norris AL, Faust J, Walter-Grimm L, Colonius T: \u003cstrong\u003eArtificial intelligence and machine learning in veterinary medicine: a regulatory perspective on current initiatives and future prospects\u003c/strong\u003e. \u003cem\u003eAm J Vet Res \u003c/em\u003e2025, \u003cstrong\u003e86\u003c/strong\u003e(S1):S16-S21.\u003c/li\u003e\n\u003cli\u003eBuono V, Sheikholharam Mashhadi P, Rahat M, Tiwari P, Byttner S: \u003cstrong\u003eExpected Grad-CAM: Towards gradient faithfulness\u003c/strong\u003e. \u003cem\u003earXiv e-prints \u003c/em\u003e2024:arXiv:2406.01274.\u003c/li\u003e\n\u003cli\u003eLa Perle KMD: \u003cstrong\u003eMachine Learning and Veterinary Pathology: Be Not Afraid!\u003c/strong\u003e \u003cem\u003eVet Pathol \u003c/em\u003e2019, \u003cstrong\u003e56\u003c/strong\u003e(4):506-507.\u003c/li\u003e\n\u003c/ol\u003e"},{"header":"Table","content":"\u003ctable border=\"0\" cellspacing=\"0\" cellpadding=\"0\" width=\"597\"\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd nowrap=\"\" colspan=\"6\" style=\"width: 597px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eTable 1. Dataset composition of canine cytological samples used in this study.\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd nowrap=\"\" valign=\"bottom\" style=\"width: 86px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd nowrap=\"\" style=\"width: 194px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd nowrap=\"\" style=\"width: 126px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eLymphoma\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd nowrap=\"\" style=\"width: 68px;\"\u003e\n \u003cp\u003e\u003cstrong\u003e\u0026nbsp;\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd nowrap=\"\" style=\"width: 54px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eRLH\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd nowrap=\"\" style=\"width: 68px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd nowrap=\"\" rowspan=\"3\" style=\"width: 86px;\"\u003e\n \u003cp\u003e# Dog\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd nowrap=\"\" style=\"width: 194px;\"\u003e\n \u003cp\u003eTraining/Validation\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd nowrap=\"\" style=\"width: 126px;\"\u003e\n \u003cp\u003e85\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd nowrap=\"\" style=\"width: 68px;\"\u003e\n \u003cp\u003e81.0%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd nowrap=\"\" style=\"width: 54px;\"\u003e\n \u003cp\u003e61\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd nowrap=\"\" style=\"width: 68px;\"\u003e\n \u003cp\u003e79.2%\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd nowrap=\"\" style=\"width: 194px;\"\u003e\n \u003cp\u003eTest\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd nowrap=\"\" style=\"width: 126px;\"\u003e\n \u003cp\u003e20\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd nowrap=\"\" style=\"width: 68px;\"\u003e\n \u003cp\u003e19.0%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd nowrap=\"\" style=\"width: 54px;\"\u003e\n \u003cp\u003e16\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd nowrap=\"\" style=\"width: 68px;\"\u003e\n \u003cp\u003e20.8%\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd nowrap=\"\" style=\"width: 194px;\"\u003e\n \u003cp\u003eTotal\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd nowrap=\"\" valign=\"bottom\" style=\"width: 126px;\"\u003e\n \u003cp\u003e105\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd nowrap=\"\" valign=\"bottom\" style=\"width: 68px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd nowrap=\"\" valign=\"bottom\" style=\"width: 54px;\"\u003e\n \u003cp\u003e77\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd nowrap=\"\" style=\"width: 68px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd nowrap=\"\" valign=\"bottom\" style=\"width: 86px;\"\u003e\u003cbr\u003e\u003c/td\u003e\n \u003ctd nowrap=\"\" style=\"width: 194px;\"\u003e\u003cbr\u003e\u003c/td\u003e\n \u003ctd nowrap=\"\" style=\"width: 126px;\"\u003e\u003cbr\u003e\u003c/td\u003e\n \u003ctd nowrap=\"\" style=\"width: 68px;\"\u003e\u003cbr\u003e\u003c/td\u003e\n \u003ctd nowrap=\"\" style=\"width: 54px;\"\u003e\u003cbr\u003e\u003c/td\u003e\n \u003ctd nowrap=\"\" style=\"width: 68px;\"\u003e\u003cbr\u003e\u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd nowrap=\"\" rowspan=\"3\" style=\"width: 86px;\"\u003e\n \u003cp\u003e# Image\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd nowrap=\"\" style=\"width: 194px;\"\u003e\n \u003cp\u003eTraining/Validation\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd nowrap=\"\" style=\"width: 126px;\"\u003e\n \u003cp\u003e1902\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd nowrap=\"\" style=\"width: 68px;\"\u003e\n \u003cp\u003e83.3%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd nowrap=\"\" style=\"width: 54px;\"\u003e\n \u003cp\u003e2493\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd nowrap=\"\" style=\"width: 68px;\"\u003e\n \u003cp\u003e80.6%\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd nowrap=\"\" style=\"width: 194px;\"\u003e\n \u003cp\u003eTest\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd nowrap=\"\" style=\"width: 126px;\"\u003e\n \u003cp\u003e382\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd nowrap=\"\" style=\"width: 68px;\"\u003e\n \u003cp\u003e16.7%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd nowrap=\"\" style=\"width: 54px;\"\u003e\n \u003cp\u003e601\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd nowrap=\"\" style=\"width: 68px;\"\u003e\n \u003cp\u003e19.4%\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd nowrap=\"\" style=\"width: 194px;\"\u003e\n \u003cp\u003eTotal\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd nowrap=\"\" valign=\"bottom\" style=\"width: 126px;\"\u003e\n \u003cp\u003e2284\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd nowrap=\"\" valign=\"bottom\" style=\"width: 68px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd nowrap=\"\" valign=\"bottom\" style=\"width: 54px;\"\u003e\n \u003cp\u003e3094\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd nowrap=\"\" valign=\"bottom\" style=\"width: 68px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n\u003c/table\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":false,"hideJournal":false,"highlight":"","institution":"","isAcceptedByJournal":false,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"veterinary-oncology","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"","sideBox":"Learn more about [Veterinary Oncology](https://veterinaryoncology.biomedcentral.com/)","snPcode":"44356","submissionUrl":"https://submission.springernature.com/new-submission/44356/3","title":"Veterinary Oncology","twitterHandle":"","acdcEnabled":true,"dfaEnabled":true,"editorialSystem":"stoa","reportingPortfolio":"BMC/SO AJ","inReviewEnabled":true,"inReviewRevisionsEnabled":true},"keywords":"Artificial intelligence, Canine lymphoma, Convolutional neural networks, Cytology, Deep learning, Explainable AI, Grad-CAM, ResNet-50","lastPublishedDoi":"10.21203/rs.3.rs-9225846/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-9225846/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003e\u003cstrong\u003eBackground\u003c/strong\u003e: Explainable artificial intelligence (AI) is a valuable tool for enhancing transparency in AI-driven image classification applications. Its practical utility and clinical evaluation remain limited in veterinary oncology, particularly within the field of diagnostic imaging. The aim of this study was to develop explainable image-based classification models using cytological data from canine lymphoma and reactive lymphoid hyperplasia (RLH).\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eMethods\u003c/strong\u003e: A total of 5,378 images were generated from cytological samples collected from 182 dogs, including 2,284 images of lymphoma (105 dogs) and 3,094 images of RLH (77 dogs). The dataset was stratified by patient and divided into training, validation, and test subsets, with model robustness evaluated through a 10-fold cross-validation protocol. Binary classification models were established using the ResNet-50 architecture. Model interpretability in a clinical context was assessed using the Gradient-weighted Class Activation Mapping (Grad-CAM) technique.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eResults:\u003c/strong\u003e The model for classifying lymphoma and RLH images achieved a test accuracy of 0.945 (95% CI: 0.929 - 0.958), a weighted precision of 0.945 (95% CI: 0.903 - 0.939), a weighted recall of 0.945 (95% CI: 0.899 - 0.969), a weighted F1-score of 0.945 (95% CI: 0.910 - 0.943), an area under the receiver operating characteristic curve of 0.979 (95% CI: 0.957 - 0.978), and an area under the precision-recall curve of 0.981 (95% CI: 0.960 - 0.985). At the patient level, 18 of 20 dogs with lymphoma were correctly classified (true positives), and all 16 dogs with RLH were correctly classified (true negatives), corresponding to a sensitivity of 0.900 (95% CI: 0.699 - 0.972), specificity of 1.000 (95% CI: 0.806 - 1.000), positive predictive value of 1.000 (95% CI: 0.824 - 1.000), and negative predictive value of 0.889 (95% CI: 0.672 - 0.969). Grad-CAM provided visual explanations of the model predictions by highlighting image regions that reflect cellular morphology pertinent to lymphoma diagnostics.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eConclusions\u003c/strong\u003e: Our findings suggest that explainable AI is applicable to cytological analysis of canine lymphoma, offering pathological insights through visual explanation of model predictions.\u003c/p\u003e","manuscriptTitle":"Applicability of explainable artificial intelligence in cytopathological evaluation of canine lymphoma","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2026-05-11 10:08:12","doi":"10.21203/rs.3.rs-9225846/v1","editorialEvents":[{"type":"communityComments","content":0},{"type":"reviewerAgreed","content":"315849515386993546732158972736555329401","date":"2026-05-05T12:02:56+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"326194837943943016062888799672618725076","date":"2026-04-16T18:49:57+00:00","index":"hide","fulltext":""},{"type":"reviewersInvited","content":"","date":"2026-04-01T17:27:08+00:00","index":"","fulltext":""},{"type":"editorAssigned","content":"","date":"2026-03-26T16:33:43+00:00","index":"","fulltext":""},{"type":"checksComplete","content":"","date":"2026-03-26T16:33:23+00:00","index":"","fulltext":""},{"type":"submitted","content":"Veterinary Oncology","date":"2026-03-25T17:03:45+00:00","index":"","fulltext":""}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"veterinary-oncology","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"","sideBox":"Learn more about [Veterinary Oncology](https://veterinaryoncology.biomedcentral.com/)","snPcode":"44356","submissionUrl":"https://submission.springernature.com/new-submission/44356/3","title":"Veterinary Oncology","twitterHandle":"","acdcEnabled":true,"dfaEnabled":true,"editorialSystem":"stoa","reportingPortfolio":"BMC/SO AJ","inReviewEnabled":true,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"ed787bcb-86b3-48cc-abaf-90994d477207","owner":[],"postedDate":"May 11th, 2026","published":true,"recentEditorialEvents":[{"type":"reviewerAgreed","content":"315849515386993546732158972736555329401","date":"2026-05-05T12:02:56+00:00","index":30,"fulltext":""}],"rejectedJournal":[],"revision":"","amendment":"","status":"under-review","subjectAreas":[],"tags":[],"updatedAt":"2026-05-11T10:08:12+00:00","versionOfRecord":[],"versionCreatedAt":"2026-05-11 10:08:12","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-9225846","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-9225846","identity":"rs-9225846","version":["v1"]},"buildId":"XKTyCvWXoU3ODBz1xrDgd","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}