INFORMATION TECHNOLOGY FOR DENTAL IMAGE SEGMENTATION BASED ON THE INTEGRATION OF LIGHTWEIGHT U-NET MODIFICATION WITH MOBILENET AND VISION MAMBA LAYERS

Abstract

Segmentation of dental images, particularly radiographs, is a crucial step in the diagnostic workflow in dentistry. High-accuracy automated segmentation methods enable more effective detection of pathologies, reduce reliance on subjective expert assessments, and accelerate clinical decision-making. The U-Net architecture remains a standard in medical image segmentation due to its capability for multiscale feature extraction and efficient utilization of skip connections. However, its large number of parameters and high computational demands pose challenges for deployment on resource-constrained devices such as mobile platforms or portable medical equipment.

This study presents a lightweight modification of U-Net aimed at improving computational efficiency without compromising segmentation accuracy. The core of the proposed model is the integration of MobileNet as the encoder, which significantly reduces the number of parameters by leveraging Depthwise Separable Convolutions. Additionally, Vision Mamba Layers are employed within the architecture to enhance the model's capacity for capturing both global and local dependencies. This combination ensures high segmentation precision while substantially reducing computational overhead.

Experimental evaluations were conducted on dental radiographs using widely recognized metrics such as Dice and IoU. The results demonstrate the advantages of the proposed model over baseline approaches, particularly in terms of segmentation accuracy, training time, and computational efficiency. The proposed architecture shows strong potential for clinical applications requiring fast and accurate dental image segmentation, as well as for deployment on devices with limited computational resources.