Leulda-Net: A Transfer Learning-Based Dual-Branch Lesion-Aware Attention Framework For Leukoplakia Detection

Abstract

Oral leukoplakia is one of the commonest potentially malignant oral lesions where early detection becomes extremely significant for avoiding progression into oral squamous cell carcinoma. Whereas manual detection process is quite subjective, laborious and dependent largely on clinical expertise, current deep learning approaches are dominated by single-backbone models, failing to capture complementary features of lesions effectively and thus being restricted to discriminate power. In order to solve these problems, this paper introduces LeuLDA-Net (Leukoplakia Lesion-Aware Dual-branch Attention Network), which is an innovative deep learning-based system for automatic detection of leukoplakia in oral images. Firstly, MSR (Multi-scale retinex) and NLM (Non-local means) filtering operations is performed on the images to increase their quality, and then data augmentation technique will be used to enrich the training set. Subsequently, the augmented images are fed into modified versions of VGG16 and Xception networks to obtain low-level and high-level lesion features, respectively. The obtained features are combined adaptively by employing Gated Feature Fusion (GFF) technique, fine-tuned through Dilated Feature Enhancement (DFE) in order to extract multi-scale context features, and further improved through Lesion-Aware Attention Module (LAM). The results of the experiment were obtained with the accuracy rate of 99.07%, the precision of 98.08%, the recall rate of 100.00%, the F1 score of 99.03%, the specificity of 98.25%, and the AUC of 0.996, which is greater than the state-of-the-art models based on deep learning techniques. Also, the application of Grad-CAM++ technique indicated that the proposed method has the capability to pay attention to the clinically important regions of lesions, thereby improving the interpretation of the results of the model.