Abstract
Oral cancer is a serious health problem worldwide and the early diagnosis of this disease can improve survival rates and prognosis of patients. In this paper, a new deep neural network, OralCancerNet, is proposed as a sequential attention-based network applicable to detecting oral lesions. The architecture builds upon a ResNet50 backbone with a Convolutional Block Attention Module (CBAM). This CBAM architecture applies the channel-wise and spatial attention successively, allowing the model to learn hierarchical feature representation through channel attention which focuses on the informative feature maps and spatial attention that attends to regions within the lesions. Strict validation on three sets of oral lesions datasets of 2,593, 750 and 1,700 images respectively reveals an impressive performance that the model has a highest accuracy of 99.42% on the primary dataset and has strong generalization on heterogeneous data sources. An ablation experiment shows that the performance of the attention mechanism in terms of the refinement of features in complex medical imaging tasks is statistically significantly better by 3.25% compared to the non-attention control condition. The model displays equal accuracy and recall in benign and malignant classes with sensitivity being particularly high in the malignancy detection. In addition to the quantitative performance, the proposed framework is transparent and can be reproduced because of its modular structure and thus, creates a solid base of future computer-aided diagnostic systems.
Keywords: Attention Mechanism, CBAM, Deep Learning, Oral Cancer Detection, Oral Lesion Recognition, OralCancerNet.