Atrous residual interconnected encoder to attention decoder framework for vertebrae segmentation via 3D volumetric CT images

Automatic vertebrae segmentation utilizing Computer Tomography (CT) plays a vital role in automated spine analyses, including the detection of vertebral body fractures and spine deformities assessment. A significant advancement in deep learning (DL) has enabled deep convolutional neural networks (DCNNs) to achieve precise performance in automated vertebrae segmentation. Despite the advantages of semantic segmentation algorithms based on DCNNs, they face limitations such as multi-scale objects, feature loss between the encoder and decoder, lack of medical image data, and limited filter field of view. A novel algorithm is presented that enables automated segmentation of vertebral bodies using volumetric CT images of the spine. The proposed model incorporates an encoder and decoder framework, and utilizes Layer Normalization to enhance the mini-batch training performance. The issue of feature loss between encoder and decoder is addressed by developing an Atrous Residual Path that carries more information from the encoder to the decoder instead of using an easy shortcut. As part of the proposed approach, a 3D Attention Module is designed to extract features from various scales in the decoding stage and further enhance the performance of the decoder. Multiple metrics are used to evaluate the proposed model on a public vertebrae dataset. According to the experimental results, our proposed approach provides competitive performance in comparison with state-of-the-art methods for automatic vertebrae semantic segmentation.