NUMERO: #20250327

Título: INCEPTIONV3-BASED DEEP LEARNING MODEL FOR HISTOLOGICAL DIFFERENTIATION OF ACINIC CELL CARCINOMA AND SECRETORY CARCINOMA

Nome do Apresentador: Sebastião Silvério SOUSA-NETO

Categoria do Trabalho: Painel de pesquisa científica (PPC)

Área Temática: Patologia Oral

Resumo: Objective: To assess the performance of the InceptionV3 convolutional neural network in differentiating acinar cell carcinoma (AciCC) from secretory carcinoma (SC) on histological slides. Study Design: A cross-sectional study was conducted using whole-slide images from 46 patients (26 AciCC, 20 SC). Slides were manually segmented into 224×224-pixel patches. The InceptionV3 model was trained and evaluated based on accuracy, sensitivity, specificity, F1-score, and AUC across training, validation, and test sets. Accuracy/loss curves and patch-level classification confidence plots were also analyzed. Results: The precision and loss curves demonstrated stable convergence and smooth learning behavior, indicating the models strong learning capacity despite dataset complexity. InceptionV3 achieved solid performance in class differentiation, supported by a final loss of 1.39 and accuracy of 0.81. Evaluation metrics included a precision of 0.73, sensitivity of 0.90, specificity of 0.73, F1-score of 0.81, and an AUC of 0.85. In the prediction certainty analysis, the model showed clear separation in class 1 predictions, reflected by a high sensitivity and a low number of false negatives. Conclusion: InceptionV3 demonstrated promising performance in differentiating AciCC and SC. Addressing current limitations-such as reliance on large datasets and the absence of clinicopathological data-could further enhance model performance.

Autor 1:  Sebastião Silvério SOUSA-NETO

E-mail 1:  [email protected]

Autor 2:   Thaís Cerqueira Reis NAKAMURA

E-mail 2:  [email protected]

Autor 3 :  Manoela Domingues MARTINS

E-mail 3:  [email protected]

Autor 4:  Fernanda Viviane MARIANO

E-mail 4:  [email protected]

Autor 5:  Anna Luiza Damaceno ARAÚJO

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]