Fine-Grain Segmentation of the Intervertebral Discs from MR Spine Images Using Deep Convolutional Neural Networks: BSU-Net

Abstract

We propose a new deep learning network capable of successfully segmenting intervertebral discs and their complex boundaries from magnetic resonance (MR) spine images. The existing U-network (U-net) is known to perform well in various segmentation tasks in medical images; however, its performance with respect to details of segmentation such as boundaries is limited by the structural limitations of a max-pooling layer that plays a key role in feature extraction process in the U-net. We designed a modified convolutional and pooling layer scheme and applied a cascaded learning method to overcome these structural limitations of the max-pooling layer of a conventional U-net. The proposed network achieved 3% higher Dice similarity coefficient (DSC) than conventional U-net for intervertebral disc segmentation (89.44% vs. 86.44%, respectively; p < 0.001). For intervertebral disc boundary segmentation, the proposed network achieved 10.46% higher DSC than conventional U-net (54.62% vs. 44.16%, respectively; p < 0.001).

Keywords: U-net; convolutional neural network; deep learning; fine grain segmentation; intervertebral disc; lumbar spine; magnetic resonance image; segmentation.

Conflict of interest statement

Conflicts of Interest: The authors declare no conflict of interest.

Figures

Figure 1.

Structure of conventional U-network (U-net).

Figure 2.

Intervertebral disc segmentation results from the conventional U-net. Blue areas are the results from the conventional U-net and red areas are manually segmented labels. Red lines are the boundaries of the labels.

Figure 3.

Whole structure of the proposed network. (a) Structure of the boundary specific U-network (BSU-net). (b) Structure of residual block. (c) Structure of BSU-pooling layer.

Figure 4.

Introduction of residual learning. (a) Conventional neural network layers. (b) A learning network of residual function S.

Figure 5.

Segmentation results of networks. (a) Dice coefficients for whole area of intervertebral discs. (b) Dice coefficients of the boundaries of intervertebral discs whose thickness is defined as 1 pixel. (c) Dice coefficients of the boundaries of intervertebral discs whose thickness is defined as 2 pixels. (d) MHDs of intervertebral discs. A paired t-test was performed to calculate p-values. * denotes p < 0.05, ** denotes p < 0.01, *** denotes p < 0.001, and n.s. denotes not significant (p > 0.05).

Figure 6.

Segmentation result from networks. Brown area, yellow area, and blue area denote correctly segmented area, under-segmented area, and over segmented area, respectively. (a) Input image. (b) U-net result. (c) Dilated U-net result. (d) BSU-net result.

Figure 7.

Segmentation results of the networks overlaid on the input image. (a) The input magnetic resonance (MR) image. (b) The input MR image with U-net segmentation result. (c) The input MR image with the result from the modified U-net which is the conventional U-net whose convolutional and pooling layers are replaced with BSU-layers. (d) The input MR image with BSU-net result.

Figure 8.

Segmentation results. (a) Input MR spine image. (b) Boundary segmentation result from U-net. (c) Boundary segmentation result from cascaded U-net. (d) Boundary segmentation result from BSU-net. White pixels correspond to boundary pixels that were perfectly matched with true boundary labels. BSU-net preserved more boundaries than other models.

Figure 9.

Segmentation results from all networks illustrating the outlier case of cascaded U-net. Brown area, yellow area, and blue area denote correctly segmented area, under-segmented area, and over segmented area, respectively. (a) Input image with label. (b) U-net result. (c) Cascaded U-net result. (d) BSU-net result.

Figure 10.

Comparison between dilated U-net and BSU-net. Blue area denotes segmentation results of dilated U-net and green area denotes segmentation results of BSU-net.