Avaliação do potencial de diferentes configurações da ResNet-50 na classificação de imagens de lâminas de tumores de glândulas salivares
Arquivos
Data
2024-09-13
Autores
Nakamura, Thaís Cerqueira Reis 
Orientadores
Moraes, Matheus Cardoso
Tipo
Trabalho de conclusão de curso
Título da Revista
ISSN da Revista
Título de Volume
Resumo
Os tumores de glândulas salivares representam um grupo heterogêneo de lesões que, se não tratadas, podem evoluir a óbito. O diagnóstico precoce e preciso é crucial para a escolha do tratamento mais adequado, promovendo uma melhor qualidade de vida para o paciente. As metodologias tradicionais de diagnóstico envolvem equipamentos de alto custo e dependem da interpretação e experiência do patologista, enquanto incertezas diminuem a precisão do diagnóstico. Em face disso, propõe-se a implementação da rede neural convolucional ResNet-50 como ferramenta de auxílio diagnóstico, classificando as lesões. Foram avaliadas quatro configurações da rede com as classes adenoma pleomórfico e carcinoma ex-adenoma pleomórfico, considerando o impacto causado pelo desequilíbrio entre as classes e a influência dos pesos inicializados por ImageNet. Os experimentos incluíram 83 pacientes, no conjunto completo foram processados 955.583 patches e para o conjunto balanceado foram fornecidos 423.420 patches na entrada da rede. Os pacientes foram divididos em conjuntos de treinamento, validação e teste para avaliar overfitting e minimizar data-leakage. O conjunto completo apresentou acurácia de 92% com ImageNet e 93% sem os pesos do ImageNet, enquanto o conjunto balanceado apresentou acurácia de 87% e 90% com e sem o uso do ImageNet, respectivamente. Todos os resultados apresentaram um bom potencial da aplicação da ResNet-50 na diferenciação das classes avaliadas. Por uma diferença sutil, os melhores resultados foram observados quando avaliado o conjunto completo sem inicialização de pesos. Os resultados mostraram muito poucas diferenças com o uso de modelos pré-treinados e balanceados. Foi observado que quando se trata de tumores tão específicos, a variabilidade tem grande valia para o resultado. Em trabalhos futuros, espera-se investigar características relacionadas à origem das amostras a fim de otimizar a separação do conjunto de treinamento. Além disso, a implementação de outras redes classificadoras se mostra promissora para esta aplicação.
Salivary gland tumors represent a heterogeneous group of lesions that, if untreated, can lead to death. Early and accurate diagnosis is crucial for selecting the most appropriate treatment, improving the patient’s quality of life. Traditional diagnostic methods involve costly equipment and rely on the pathologist's interpretation, while uncertainties reduce diagnostic accuracy. Therefore, the implementation of the ResNet-50 convolutional neural network is proposed as a diagnostic tool to classify these lesions. Evaluations were conducted on network configurations with four setups, with the classes of pleomorphic adenoma and carcinoma ex-pleomorphic adenoma, considering class imbalance and the impact of weights initialized by ImageNet. A total of 83 patients were evaluated in all experiments, with 955,583 patches processed for the complete set and 423,420 patches for the balanced set. Patients were divided into training, validation, and test sets to assess overfitting and minimize data-leakage. The complete set showed 92% accuracy with ImageNet weights and 93% without, while the balanced set showed 87% and 90% accuracy with and without ImageNet weights, respectively. All results demonstrated good potential for applying ResNet-50 in differentiating the evaluated classes. The best results were subtly observed with the complete set without weight initialization. Results indicated minimal differences when using pre-trained and balanced models. It was observed that when dealing with such specific tumors, variability greatly affects outcomes. Future work will investigate specific characteristics related to sample origin to optimize the training set separation. Additionally, implementing other classifier networks appears promising for this application.
Salivary gland tumors represent a heterogeneous group of lesions that, if untreated, can lead to death. Early and accurate diagnosis is crucial for selecting the most appropriate treatment, improving the patient’s quality of life. Traditional diagnostic methods involve costly equipment and rely on the pathologist's interpretation, while uncertainties reduce diagnostic accuracy. Therefore, the implementation of the ResNet-50 convolutional neural network is proposed as a diagnostic tool to classify these lesions. Evaluations were conducted on network configurations with four setups, with the classes of pleomorphic adenoma and carcinoma ex-pleomorphic adenoma, considering class imbalance and the impact of weights initialized by ImageNet. A total of 83 patients were evaluated in all experiments, with 955,583 patches processed for the complete set and 423,420 patches for the balanced set. Patients were divided into training, validation, and test sets to assess overfitting and minimize data-leakage. The complete set showed 92% accuracy with ImageNet weights and 93% without, while the balanced set showed 87% and 90% accuracy with and without ImageNet weights, respectively. All results demonstrated good potential for applying ResNet-50 in differentiating the evaluated classes. The best results were subtly observed with the complete set without weight initialization. Results indicated minimal differences when using pre-trained and balanced models. It was observed that when dealing with such specific tumors, variability greatly affects outcomes. Future work will investigate specific characteristics related to sample origin to optimize the training set separation. Additionally, implementing other classifier networks appears promising for this application.
Descrição
Citação
NAKAMURA, T. C. R. Avaliação do Potencial de Diferentes Configurações da ResNet-50 na Classificação de Imagens de Lâminas de Tumores de Glândulas Salivares. 2024. 11 f. Trabalho de Conclusão de Curso - Instituto de Ciência e Tecnologia da Universidade Federal de São Paulo, São José dos Campos, 2024.