The use of deep learning state-of-the-art architectures for oral epithelial dysplasia grading: A comparative appraisal

• Published in: Journal of oral pathology & medicine Vol. 52; no. 10; pp. 980 - 987
• Main Authors: Araújo, Anna Luíza Damaceno, Silva, Viviane Mariano da, Moraes, Matheus Cardoso, de Amorim, Henrique Alves, Fonseca, Felipe Paiva, Sant'Ana, Maria Sissa Pereira, Mesquita, Ricardo Alves, Mariz, Bruno Augusto Linhares Almeida, Pontes, Hélder Antônio Rebelo, de Souza, Lucas Lacerda, Saldivia-Siracusa, Cristina, Khurram, Syed Ali, Pearson, Alexander T, Martins, Manoela Domingues, Lopes, Marcio Ajudarte, Vargas, Pablo Agustin, Kowalski, Luiz Paulo, Santos-Silva, Alan Roger
• Format: Journal Article
• Language: English
• Published: Denmark Wiley Subscription Services, Inc 01-11-2023
• Subjects: Biopsy; Cross-Sectional Studies; Deep Learning; Dysplasia; Humans; Machine Learning; Neural networks; Neural Networks, Computer; dysplasia grading; leukoplakia; artificial intelligence; erythroleukoplakia; whole slide images
• ISSN: 0904-2512; 1600-0714; 1600-0714
• DOI: 10.1111/jop.13477

Dysplasia grading systems for oral epithelial dysplasia are a source of disagreement among pathologists. Therefore, machine learning approaches are being developed to mitigate this issue.
This cross-sectional study included a cohort of 82 patients with oral potentially malignant disorders and correspondent 98 hematoxylin and eosin-stained whole slide images with biopsied-proven dysplasia. All whole-slide images were manually annotated based on the binary system for oral epithelial dysplasia. The annotated regions of interest were segmented and fragmented into small patches and non-randomly sampled into training/validation and test subsets. The training/validation data were color augmented, resulting in a total of 81,786 patches for training. The held-out independent test set enrolled a total of 4,486 patches. Seven state-of-the-art convolutional neural networks were trained, validated, and tested with the same dataset.
The models presented a high learning rate, yet very low generalization potential. At the model development, VGG16 performed the best, but with massive overfitting. In the test set, VGG16 presented the best accuracy, sensitivity, specificity, and area under the curve (62%, 62%, 66%, and 65%, respectively), associated with the higher loss among all Convolutional Neural Networks (CNNs) tested. EfficientB0 has comparable metrics and the lowest loss among all convolutional neural networks, being a great candidate for further studies.
The models were not able to generalize enough to be applied in real-life datasets due to an overlapping of features between the two classes (i.e., high risk and low risk of malignization).