Authors:
Marcelo Nogueira
1
;
2
and
Elsa Gomes
3
;
1
Affiliations:
1
INESC TEC, Porto, Portugal
;
2
Faculdade de Ciências da Universidade do Porto, Departamento de Ciência de Computares, Porto, Portugal
;
3
Instituto Superior de Engenharia do Porto, Porto, Portugal
Keyword(s):
Oral Cancer, Histopathology, Deep Learning, CNN, Image Classification, Transfer Learning, Data Augmentation, Data Leakage.
Abstract:
Oral squamous cell carcinoma is one of the most prevalent and lethal types of cancer, accounting for approximately 95% of oral cancer cases. Early diagnosis increases patient survival rates and has traditionally been performed through the analysis of histopathological images by healthcare professionals. Given the importance of this topic, there is an extensive body of literature on it. However, during our bibliographic research, we identified clear cases of data leakage related to contamination of test data due to the improper use of data augmentation techniques. This impacts the published results and explains the high accuracy values reported in some studies. In this paper, we evaluate several models, with a particular focus on EfficientNetBx architectures combined with Transformer layers, which were trained using Transfer Learning and Data Augmentation to enhance the model’s feature extraction and attention capabilities. The best result, obtained with the Effi-cientNetB0, together
with the Transformer layers, achieved an accuracy rate of 87.1% on the test set. To ensure a fair comparison of results, we selected studies that we identified as not having committed data leakage.
(More)