NUMERO: #20240029

Título: ADDRESSING HISTOLOGICAL OVERLAP BETWEEN PLEOMORPHIC ADENOMA AND CARCINOMA EX-PLEOMORPHIC ADENOMA THROUGH DEEP LEARNING

Nome do Apresentador: Sebastião Silvério de Sousa NETO

Categoria do Trabalho: Painel de pesquisa científica (PPC)

Área Temática: Patologia Oral

Resumo: Objective: To develop a model capable of distinguishing carcinoma ex-pleomorphic adenoma from pleomorphic adenoma using a Convolutional Neural Network architecture.
Study design: A cohort of 30 patients, equally divided into carcinoma ex-pleomorphic adenoma (n=15) and pleomorphic adenoma (n=15), was used for training a Convolutional Neural Network. The whole-slide images were annotated and fragmented into patches of 299 x 299 pixels. A total of 169,544 (carcinoma ex-pleomorphic adenoma) and 112,320 (pleomorphic adenoma) patches were generated. Training (80%), validation (10%), and test (10%) subsets were established. The Residual Neural Network (ResNet)-50 was chosen for its recognition and classification capabilities. We assessed the model by generating graphs and extracting parameters from both the training and validation sets. Additionally, we computed the accuracy index using the confusion matrix.
Results: The developed model demonstrated learning potential but exhibited limitations in generalization, as indicated by the validation curve. The loss curve recorded 7.52, and the accuracy reached 0.56. Evaluated parameters included specificity (0.76), sensitivity (0.36), precision (0.69), F1 score (0.48), and AUC (0.69).
Conclusion: Although there are limitations on generalization, the highlighted good specificity indicates potential to evolve. Future efforts will focus on developing innovative deep learning techniques, integrating larger datasets, and encouraging inter-institutional collaboration.

Autor 1:  Sebastião Silvério de Sousa NETO

E-mail 1:  [email protected]

Autor 2:   Anna Luiza Damaceno ARAÚJO

E-mail 2:  [email protected]

Autor 3 :  Giovanna Calabrese dos SANTOS

E-mail 3:  [email protected]

Autor 4:  Manoela Domingues MARTINS

E-mail 4:  [email protected]

Autor 5:  Alan Roger SANTOS-SILVA

E-mail 5:  [email protected]

Autor 6:  Matheus Cardoso MORAES

E-mail 6:  [email protected]

Autor 7:  Pablo Agustin VARGAS

E-mail 7:  [email protected]