An Enhanced Caries Detection and Prediction Using DentSU_Net

Segmenting medical images is crucial for clinical diagnosis and case analysis. Currently, most of the successful techniques are based on U-shaped encode and decoder-based convolutional neural networks (CNNs). A major drawback of these approaches is their limited capacity to establish highly relevant pattern connections and comprehensive contextual associations, leading to imprecisions during segmenting regions of interest. To address this limitation and improve the utilization of global semantic features with minimal semantic representation between encoding and decoding stages, the proposed work DentSU_Net enhanced independent channel elite features using Swin_Unet incorporated through Squeeze and Excitation (SE) attention layer to effectively segment the dental caries from panoramic X-ray images. This layer selectively emphasizes informative features and suppresses less relevant features recalibrating the importance of different channels for better learning and representation features for precisely identifying the caries region from dental images. The experimental results state that the proposed DentSU_Net attained a dice coefficient of 0.90, accuracy of 0.97, IoU of 0.83, sensitivity of 0.85, and specificity of 0.99. The results outperform the performance of prior studies and can help dentists in automatically and more effectively segmenting the caries zone.