Caracterização de imagens de tecido mamário felino corado com Picrosirius Red

Abstract

Feline mammary carcinoma represents the third most frequent tumor in queens. It exhibits aggressive behavior and a high mortality rate due to its invasive and metastatic pattern, making it a relevant model for comparative studies with breast cancer. Stromal alterations, such as increased collagen deposition, are associated with tumor aggressiveness and can be evaluated using histological techniques such as Picrosirius Red (PSR), complementing the routine Hematoxylin–Eosin (H&E) staining used for histopathological diagnosis. Computational methods, including feature extraction and artificial intelligence, enable image analysis to improve diagnostic accuracy. In this context, the present study aimed to classify images of healthy and tumoral feline mammary tissue stained with PSR according to histopathological grading. To this end, methods involving collagen quantification in the images were employed, as well as image characterization using Haralick texture descriptors and statistical analysis, with Gaussian curves applied for feature selection. Finally, machine learning models based on Random Forest (RF), XGBoost (XGB), and Multilayer Perceptron (MLP) were used with the selected features. Seven features were selected to classify healthy tissue and tumor images, and five features were selected to distinguish low- and high-grade tumors. The MLP model achieved the best performance in classifying healthy tissue (26 of 28 images) and tumor tissue (30 of 30 images), reaching 96.4% accuracy on the test set, whereas XGB demonstrated superior performance in differentiating low-grade (30 of 38 images) and high-grade tumors (36 of 27 images), achieving 88.0% accuracy. Skewness and kurtosis contributed to the classification between healthy and altered tissue images but were not effective in distinguishing low- and high-grade carcinogenic tissue. A subset of Haralick descriptors contributed to both classification tasks, mainly due to their ability to capture the heterogeneity of tumor tissues regarding collagen deposition during carcinogenesis, in contrast to the homogeneity of thinner collagen fibers in healthy feline mammary tissue. These results reinforce the potential of texture analysis combined with artificial intelligence methods as a complementary tool to histopathological evaluation and for improving diagnostic accuracy in veterinary oncology. With further refinements, the same approach may also assist in the diagnosis of breast cancer.