Assessment of the association of deep features with a polynomial algorithm for automated oral epithelial dysplasia grading

• Main Authors: Silva, Adriano B., De Oliveira, Cleber I., Pereira, Danilo C., Tosta, Thaina A. A., Martins, Alessandro S., Loyola, Adriano M., Cardoso, Sergio V., De Faria, Paulo R., Neves, Leandro A., Do Nascimento, Marcelo Z.
• Format: Conference Proceeding
• Language: English
• Subjects: Classification algorithms; Feature extraction; Histopathology; Lesions; Mice; Tongue; Transfer learning
• ISSN: 2377-5416
• DOI: 10.1109/SIBGRAPI55357.2022.9991758

Oral epithelial dysplasia is a potentially malignant lesion that presents challenges for diagnosis. The use of digital systems in histological analysis can aid specialists to obtain data that allows a robust and fast grading process, but there are few methods in the literature proposing a grading system for this lesion. This study presents a method for oral epithelial dysplasia grading in histopathological images combining deep features and a polynomial classifier. The ResNet50 and AlexNet models were trained with the images and information was extracted from the convolutional layers, exploring convolutional neural networks via transfer learning. Then, the ReliefF algorithm was used to rank and select the most relevant features, which were given as an input to the polynomial classifier. The methodology was employed in a dataset with 296 regions of mice tongue images. The results were compared with the gold standard and other algorithms present in the literature. The classification stage presented AUC values ranging from 0.9663 to 0.9800. When compared to other algorithms present in the literature, our method provided relevant results regarding accuracy and AUC values. The proposed approach presented relevant results and can be used as a tool to aid pathologists in grading oral dysplastic lesions.