Attention U-Net Architecture for Dental Image Segmentation

Abstract

Dental image segmentation is a very challenging task due to the presence of multiple teeth in the X-Ray and due to its different size in each human being. U-Net Architecture has been used in the Pancreas Segmentation [1] and Lung Segmentation [2] but the feature map of dental image is different from that of the lungs and pancreas. Technology used in the screening of the lungs and pancreas is CT Scan which is different form the standard X-Ray. To adapt the dental image feature we have proposed a U-Net Architecture [3] along with the attention gate [4]. U-Net Architecture [3] created by Olaf Ronneberger, Thomas Brox and Philipp Fischer is generally used in biomedical segmentation of images. It uses encoder and decoder architecture for the image segmentation but still there may be information losses while encoding the layers for feature map and unnecessary information is passed using the skip connection while decoding in standard U-Net Architecture. Limitation of the plain skip connection can be solved using the Attention block. Using Attention block in the U-Net model will further increase the sensitivity and prediction accuracy with minimum computational overhead. Model trained with attention block self-learn themselves. Any irrelevant regions in the input image are suppressed, and hence only the important features are highlighted. In our network architecture we have achieved the accuracy of 92.8 percent and 94.7 percent dice score.