Enhancing Explainability in Oral Cancer Detection with Grad-CAM Visualizations

Arnaldo Vitor Barros da Silva, Cristina Saldivia-Siracusa, Eduardo Santos Carlos de Souza, Anna Luíza Damaceno Araújo, Marcio Ajudarte Lopes, Pablo Agustin Vargas, Luiz Paulo Kowalski, Alan Roger Santos-Silva, André C. P. L. F. de Carvalho, Marcos G. Quiles

Abstract: Late diagnosis of oral cancer significantly compromises patient outcomes. A promising approach to speed up the diagnostic process involves the use of Deep Learning (DL) models for medical image analysis. However, a notable challenge with these models is their lack of interpretability. To address this, techniques like Gradient-weighted Class Activation Mapping (Grad-CAM) have been developed. Grad-CAM generates heatmaps that highlight image regions most influential for classification decisions. In our study, we evaluated the performance of two DL models renowned for their high accuracy in oral cancer classification. Our analysis extended beyond mere accuracy metrics; we employed Grad-CAM to provide visual explanations of the models’ decisions. Furthermore, we investigated subclass accuracy rates and the distribution of prediction confidences to gain a deeper insight into the models’ performance and robustness in oral cancer detection. This comprehensive evaluation approach offers a more nuanced understanding of the capabilities and limitations of DL methods in the context of oral cancer diagnosis.