Graduação de mastocitomas cutâneos caninos por citopatologia: previsão de aprendizagem de máquina
Guardat en:
| Autor principal: | |
|---|---|
| Data de publicació: | 2024 |
| Format: | Master thesis |
| Idioma: | por |
| Font: | Repositório Institucional da UPF |
| Download full: | https://repositorio.upf.br/handle/123456789/10039 |
Sumari: | Mast cell tumor is a malignant tumor caused by the proliferation of mast cells, with a high prevalence, accounting for 10% of cutaneous tumors in dogs. The diagnosis of this neoplasm can be performed through cytopathology as a screening method. Mathematical models aim to describe, explain, and predict phenomena and processes, serving as an important tool that can aid in understanding biological systems that cannot be fully comprehended through verbal reasoning alone. Camus et al. (2016) proposed a cytological classification system in which mast cell tumor is considered high-grade if there is poor granulation or two of four other characteristics (presence of any mitotic figures, pleomorphism, binucleation or multinucleation, or marked anisokaryosis [> 50% variation in nuclear size]). Although the developed algorithm is used in laboratory routines and has high specificity and sensitivity for high-grade tumors, there remains a possibility for onethird of low-grade tumors to be incorrectly classified as high-grade in cytopathological examination. Therefore, other models for the classification of mast cell tumors using cytological examination are needed to achieve a desirable level of confidence. Thus, the present study aimed to obtain a mathematical model with greater accuracy for determining mast cell tumors. For this purpose, 99 dogs of different breeds and both sexes were evaluated, undergoing cytopathological and histopathological examinations with a diagnosis of cutaneous mast cell tumor. Inclusion parameters were established, such as the cytopathological description of specific criteria (pleomorphism, anisocytosis, anisokaryosis, number of nucleoli and nucleolar alterations, mesenchymal involvement, eosinophilic infiltration, granulation level, inflammation grade, multinucleation, and presence of mitotic figures), in a standardized manner, with histopathological examination performed within a time interval of up to eight weeks, using the Kiupel system as the outcome variable. Based on case selection, the dataset was randomly divided into two parts, with 70% of observations randomly assigned to the training and validation set and the remaining 30% to evaluate performance. For model training, a generalized linear model was used (binomial family, logit link function). Given the limited size of the dataset, the k-fold cross-validation technique was used for its ability to produce unbiased performance results, opting for a 10-fold cross-validation strategy. Subsequently, the model’s predictive values were calculated from the test dataset, and performance metrics such as accuracy, kappa, and no-information rate (NIR) were evaluated using the "confusionMatrix()" function from the "caret" package. The model achieved an accuracy of 0.86, with a 95% confidence interval ranging from 0.66 to 0.96, and a p-value of 0.012, with a sensitivity of 0.80 and a specificity of 0.89. The results obtained were able to reduce the probability of classifying low-grade mast cell tumors as high-grade, compared to previous studies. Furthermore, the variables identified as important for model classification were “maximum number of nucleoli,” “mitotic figures,” “anisokaryosis,” and “anisocytosis.” |