LCSNet: Lightweight Caries Segmentation Network for the segmentation of dental caries using smartphone photographs

Dental caries is one of the major dental issues that is common among many individuals. It leads to tooth loss and affects the tooth root, creating a need to automatically detect dental caries to reduce treatment costs and prevent its consequences. The Lightweight Caries Segmentation Network (LCSNet) proposed in this study detects the location of dental caries by applying pixel-wise segmentation to dental photographs taken with various Android phones. LCSNet utilizes a Dual Multiscale Residual (DMR) block in both the encoder and decoder, adapts transfer learning through a pre-trained InceptionV3 model at the bottleneck layer, and incorporates a Squeeze and Excitation block in the skip connection, effectively extracting spatial information even from images where 95% of the background and only 5% represent the area of interest. A new dataset was developed by gathering oral photographs of dental caries from two hospitals, with advanced augmentation techniques applied. The LCSNet architecture demonstrated an accuracy of 97.36%, precision of 73.1%, recall of 70.2%, an F1-Score of 71.14%, and an Intersection-over-Union (IoU) of 56.8%. Expert dentists confirmed that the LCSNet model proposed in this in vivo study accurately segments the position and texture of dental caries. Both qualitative and quantitative performance analyses, along with comparative analyses of efficiency and computational requirements, were conducted with other deep learning models. The proposed model outperforms existing deep learning models and shows significant potential for integration into a smartphone application-based oral disease detection system, potentially replacing some conventional clinically adapted methods. • A new dataset is prepared using different company smartphone cameras with variations in resolution, illumination, and texture for the pixel-level segmentation of dental caries. • A novel LCSNet deep learning architecture was proposed. • A performance comparison with state-of-the-art methods was presented. • LCSNet trained with 5.87 million parameters and achieved a good F1-score(71.14%) and IoU of 56.8%, and is computationally efficient as it takes less inference time (3131.75 milliseconds) and fewer floating-point operations (FLOPs)(31.58 billion) compared to other segmentation models. • Expert dentists from two of the hospitals validated the results.