LumNet: A Deep Neural Network for Lumbar Paraspinal Muscles Segmentation

Abstract

Lumber paraspinal muscles(LPM) segmentation is of essential importance in predicting response to treatment of low back pain. To date, all LPM segmentation methods are manually based instead of automatic. Manual segmentation of LPM requires vast radiological knowledge and experience. Moreover, the manual segmentation usually induces subjective variance. Therefore, an automatic segmentation is desireable. It is challenging to achieve automatic segmentation mainly because the ambiguous boundary of the LPM can be very difficult to locate. In this paper, we present a novel encoder-decoder and attention based deep convolutional neural network (CNN) to address this problem. With the help of skip connections, the encoder-decoder structure can capture both shadow and deep features which represent local and global information. Pre-trained VGG11 in ImageNet performed as encoder. In the decoder part, an attention block is applied to recalibrate the input feature. With the help of attention block, meaningful features are highlighted while irrelevant features are suppressed. To fully evaluate the performance of our proposed network, we construct the first large-scale LPM segmentation dataset with 1080 images and its segmentation masks. Experimental results show that our proposed network can not only achieve a good LPM segmentation result with a high dice score of 0.94 but also outperforms other state-of-the-art segmentation methods.

An externally validated deep learning model for the accurate segmentation of the lumbar paravertebral muscles

“…Other automated solutions based on neural networks aimed at segmenting the lumbar paravertebral muscles have been previously presented. Li et al and achieved similar performances [20]. However, it should be noted that in both studies psoas major and quadratus lumborum were not segmented, and that an internal validation on a test set randomly selected from the original database rather than an external validation of a distinct dataset was performed.…”

“…Li et al [12] developed a U-Net based network to perform the segmentation of multifidus and erector spinae, obtaining average Dice similarity coefficients of 0.949 and 0.913 respectively, generally in line with the performance of the current model while taking into account that the metrics used in the present and in the literature study are not directly comparable. Zhang et al developed another U-Net based model and achieved similar performances [20]. However, it should be noted that in both studies psoas major and quadratus lumborum were not segmented, and that an internal validation on a test set randomly selected from the original database rather than an external validation of a distinct dataset was performed.…”

Advancing spine care through AI and machine learning: overview and applications

“…Automatic tissue segmentation from MRI data showcases the ability of machine learning to perform complex tasks, such as vertebrae recognition and segmentation with a high degree of accuracy, comparable to manual segmentation by physicians (56,57,58). Attention has also been devoted to the segmentation of paravertebral muscles (59,60,61), the properties of which (such as cross-sectional areas and fatty infiltration) are known to be associated with disability and treatment outcomes (62).…”