Feasibility of deep learning algorithm in diagnosing lumbar central canal stenosis using abdominal CT

Abstract

Objective: To develop a deep learning algorithm for diagnosing lumbar central canal stenosis (LCCS) using abdominal CT (ACT) and lumbar spine CT (LCT).

Materials and methods: This retrospective study involved 109 patients undergoing LCTs and ACTs between January 2014 and July 2021. The dural sac on CT images was manually segmented and classified as normal or stenosed (dural sac cross-sectional area ≥ 100 mm² or < 100 mm², respectively). A deep learning model based on U-Net architecture was developed to automatically segment the dural sac and classify the central canal stenosis. The classification performance of the model was compared on a testing set (990 images from 9 patients). The accuracy, sensitivity, and specificity of automatic segmentation were quantitatively evaluated by comparing its Dice similarity coefficient (DSC) and intraclass correlation coefficient (ICC) with those of manual segmentation.

Results: In total, 990 CT images from nine patients (mean age ± standard deviation, 77 ± 7 years; six men) were evaluated. The algorithm achieved high segmentation performance with a DSC of 0.85 ± 0.10 and ICC of 0.82 (95% confidence interval [CI]: 0.80,0.85). The ICC between ACTs and LCTs on the deep learning algorithm was 0.89 (95%CI: 0.87,0.91). The accuracy of the algorithm in diagnosing LCCS with dichotomous classification was 84%(95%CI: 0.82,0.86). In dataset analysis, the accuracy of ACTs and LCTs was 85%(95%CI: 0.82,0.88) and 83%(95%CI: 0.79,0.86), respectively. The model showed better accuracy for ACT than LCT.

Conclusion: The deep learning algorithm automatically diagnosed LCCS on LCTs and ACTs. ACT had a diagnostic performance for LCCS comparable to that of LCT.

Keywords: Artificial Intelligence; Central canal stenosis; Computed tomography; Deep learning; Lumbar spine.

Fig. 1

Flowchart of datasets for development and test. L-spine CT = Lumbar spine CT, AP CT = Abdominal CT

Fig. 2

Development and test datasets. In the development set (a), one disc level for each patient was randomly assigned to L1-L2, L2-L3, L3-L4, L4-L5, or L5-S1. In the test set (b), all disc levels for each patient were included. For each level, a total of 11 slices were taken from the center of the intervertebral disc, five slices above and below the center with an interval of 1 mm, constituting a total of 2,200 images for the development set and 990 images for the test set. The dotted red line shows the center of disc space and the orange dot the above and below the center of disc space with an interval of 1 mm

Fig. 3

Representative images of the labeling program. The program automatically detects the bone of the lumbar spine (indicated by blue color in the image). Labeling of the dural sac area (pink), disc posterior margin (light pink), and ligamentum flavum (green) on an axial CT scan by the radiologists and students (a). Automatic reformatting of sagittal (b) image by the program

Fig. 4

Overview of the proposed network architecture. The U-Net architecture was used as the model architecture. Each box corresponds to a multi-channel capability map. Convolution is a layer that applies a convolutional filter to the image and feature map to compress the information efficiently. Batch normalization is a technique that helps different features to have a consistent distribution, and ReLU is an activation function that maps the outputs of the layer to values greater than or equal to zero so that errors are evenly transmitted during the learning process. Max pooling is a method of generating a smaller feature map by taking the maximum value in each region of the feature map, and Up-convolution is a layer that increases the size by applying a convolution filter to the feature map. Concatenation stacks feature maps of the same size for a channel dimension

Fig. 5

Representative cases of dural sac segmentation on a normal case of lumbar CT without contrast. The images of axial CT imaging (a) selected from the dataset are shown alongside their resulting segmentation of the dural sac by a manual rater (b, red) and the proposed algorithm (c, green). The dural sac area is 127.6 mm² (b) and 124.2 mm² (c), respectively. Figure 5d is an MR axial image of this case, which was taken within 6 months of the CT scan

Fig. 6

Representative cases of dural sac segmentation on a stenosis case with contrast on abdominal CT. The images of axial enhanced abdominal CT imaging (a) selected from the dataset are shown alongside their resulting segmentation of the dural sac by a manual rater (b, red) and the proposed algorithm (c, green). The dural sac area is 50.6 mm² (b) and 45.6 mm² (c), respectively. Figure 6d is an MR axial image of this case, which was taken within 6 months of the CT scan