A Novel Feature Extraction Approach Using Deformable Adaptive Instance-Based U-Net Architecture for Segmentation and Classification of Oral Mucosal Lesion
S. M. Sagari, Vindhya P. Malagi, B. Chandrahas · Engineering Technology & Applied Science Research · 2025
Oral cancer is one of the six cancer types having high morbidity and mortality rates, especially among socioeconomically deprived groups of people due to their lack of knowledge about oral hygiene. This study aimed to detect oral lesions in different areas of the oral cavity based on visual features of suspicious regions. The localization, detection, and classification of regions of interest in digital images stemming from diverse resolution cameras presents a formidable challenge due to the variation in illumination, image size, and varied noise. The proposed method employs image pre- and post-processing approaches to locate the regions effectively. A dataset of 2050 oral cavity images was used, having 1000 malignant, 700 benign, and 350 premalignant cases. The proposed method uses deformable convolution and instance normalization in the U-Net architecture to segment the region of interest by preprocessing the images using canny and local binary pattern feature extractors. These segmented regions are classified by combining the Bresenham circle and flood fill algorithms. The experimental analysis of the proposed approach showed precision, recall, and F1 scores of 93.85%, 97.37%, and 95.58% for noised malignant images and 96.20%, 96.82%, and 96.51% for denoised malignant images. Similarly, precision, recall, and F1 scores were 98.67%, 94.94%, and 96.77% for benign lesion noise images, and 96.95%, 96.36%, and 96.66% for benign denoised lesion images.