DEEP LEARNING FOR SALIVARY GLAND TUMOR CLASSIFICATION.

The present study proposes to develop and implement a Deep Learning model for automatic classification of clinical photographs of salivary gland tumors in the palate into benign and malignant categories. A dataset of 100 clinical images of SGT from seven institutions was used to train and validate a ResNet50 (original architecture) implemented with a low learning rate of 10-5 for 75 epochs with 10-fold cross-validation. The proposed ResNet50 reached an accuracy of 70% and an AUC of 0.68 during training, showing the potential for learning. However, divergence in training and validation accuracy and loss curves displayed a clear overfitting, which is not uncommon when training Deep Learning algorithms with a small sample. The proposed DL-based model presented a capacity for learning with the potential of achieving a fair accuracy. To overcome overfitting and improve the results, further steps of the present investigation will consider transfer learning and data augmentation. [ABSTRACT FROM AUTHOR]