Histopathologic Analysis of Canine Tumors Using Convolutional Neural Networks

Douglas Rodrigues, Joao Paulo Papa, Rebeca Scalco, FAVIANA AIMÉ MAZA LEÓN, Alessandre Hataka, Clayton Pereira

Track: Track 6: Applications of Intelligent Systems in Complex Environments

Keywords: Deep Learning, Explainable artificial intelligence (XAI), Histopathologic Analysis, Veterinary Pathology, Mast Cell Tumors, Squamous Cell Carcinoma

TL;DR: Explainability analyses showed that the models focused on relevant histological regions, indicating biologically plausible decision-making.

Abstract: Veterinary pathology plays a fundamental role in clinical decision-making, with histopathological examination serving as the gold standard for diagnosing canine neoplasms. Among these, mast cell tumors and squamous cell carcinomas are highly prevalent, and their accurate differentiation is essential for prognosis and treatment planning. However, traditional grading workflows remain subjective, labor-intensive, and susceptible to interobserver variability. This study investigates deep learn- ing–based approaches for automated classification of canine mast cell tumors and squamous cell carcinomas using a curated dataset of hematoxylin-and-eosin (H&E) stained images. Thirteen state-of-the-art Convolutional Neural Network architectures were systematically evaluated under two learning-rate configurations to assess the influence of network depth, connectivity patterns, and optimization hyperparameters on model performance. The results show that Xception and ResNet-152 achieved the best performance at lr = 0.001, whereas InceptionResNetV2 and DenseNet-121 attained the highest accuracies and perfect ROC–AUC scores at lr = 0.0001. These findings highlight that both architectural choice and learning-rate selection critically affect convergence stability and predictive accuracy. Explainability analyses based on Grad-CAM and feature activation visualization confirmed that the models focused on histologically meaningful regions, supporting the biological plausibility of their decision processes. Overall, this study demonstrates the potential of deep learning to enhance diagnostic consistency, objectivity, and scalability in veterinary pathology, paving the way for more reliable computational support tools in clinical workflows.

Submission Number: 15