NUMERO: #20250373

Título: DEEP LEARNING APPROACHES FOR DIFFERENTIATING LIP CANCER FROM ACTININC CHEILITIS: A MODEL COMPARISON STUDY

Nome do Apresentador: Ivan José Correia NETO

Categoria do Trabalho: Painel de pesquisa científica (PPC)

Área Temática: Estomatologia

Resumo: Objective: This study evaluated six deep learning models (EfficientNet B7, InceptionResNetV2, MobileNetV2, ResNet-152, U-Net MobileNetV2, and VGG16) trained to distinguish clinical images of actinic cheilitis (n=397) from lip squamous cell carcinoma (n=92), totaling 489 images. Methods: The dataset was split into training (80%), validation (10%), and test (10%) sets, preserving class proportions and avoiding data leakage. Two training strategies were tested: (1) a baseline model without balancing; and (2) a model with undersampling and data augmentation. Transfer learning was applied using ImageNet-pretrained weights with fine-tuning. Results: In the baseline, VGG16 performed best (accuracy: 89.6%; F1-score: 0.88; precision: 1.0), despite class imbalance. After balancing, InceptionResNetV2 achieved the highest AUC-ROC (0.866), and VGG16 showed better sensitivity. Conclusion: These findings highlight the importance of balancing and regularization techniques in improving model performance, particularly under clinical conditions marked by lesion heterogeneity and class imbalance.

Autor 1:  Ivan José Correia NETO

E-mail 1:  [email protected]

Autor 2:   Ayrton da Costa GANEM FILHO

E-mail 2:  [email protected]>

Autor 3 :  Anna Luíza Damaceno ARAÚJO

E-mail 3:  [email protected]

Autor 4:  Luiz Paulo KOWALSKI

E-mail 4:  [email protected]

Autor 5:  Alan Roger SANTOS-SILVA

E-mail 5:  [email protected]

Autor 6:  André Carlos Ponce de Leon Ferreira de CARVALHO

E-mail 6:  [email protected]

Autor 7:  Márcio Ajudarte LOPES

E-mail 7:  [email protected]