Boundary Delineation of MRI Images for Lumbar Spinal Stenosis Detection Through Semantic Segmentation Using Deep Neural Networks

Abstract

We propose a methodology to aid clinicians in performing lumbar spinal stenosis detection through semantic segmentation and delineation of magnetic resonance imaging (MRI) scans of the lumbar spine using deep learning. Our dataset contains MRI studies of 515 patients with symptomatic back pains. Each study is annotated by expert radiologists with notes regarding the observed characteristics and condition of the lumbar spine. We have developed a ground truth dataset, containing image labels of four important regions in the lumbar spine, to be used as the training and test images to develop classification models for segmentation. We developed two novel metrics, namely confidence, and consistency, to assess the quality of the ground truth dataset through a derivation of the Jaccard Index. We experimented with semantic segmentation of our dataset using SegNet. Our evaluation of the segmentation and the delineation results show that our proposed methodology produces a very good performance as measured by several contourbased and region-based metrics. In addition, using the Cohen's kappa and frequency-weighted confidence metrics, we can show that 1) the model's performance is within the range of the worst and the best manual labeling results and 2) the ground-truth dataset has an excellent inter-rater agreement score. We also presented two representative delineation results of the worst and best segmentation based on their BF-score to show visually how accurate and suitable the results are for computer-aided-diagnosis purposes.

Automated measurement of anteroposterior diameter and foraminal widths in MRI images for lumbar spinal stenosis diagnosis

“…We adopted the transfer learning approach [ 31 ] when training the SegNet model instead of developing the model from scratch since we have shown previously that the former approach produces significantly better segmentation than the latter approach [ 27 ]. We did this by using a SegNet model with pre-trained VGG16 coefficients on the ImageNet database [ 32 ].…”

“…We have previously shown the applicability of the SegNet architecture in segmenting axial MRI images using unmodified label images [ 27 ]. We will use the previously reported results as a baseline to measure the improvement in the segmentation accuracy along the important boundaries using the proposed method.…”

“…Even when the process is done by a skilled labeler, there is still a high level of probabilities of inaccuracies, especially along region boundaries. To illustrate this point, we show in Fig 5 an example of the region boundaries of a manually created label image and its corresponding automatically segmented image using the SegNet model we developed previously in [ 27 ]. Fig 5A clearly shows the challenges in manual segmentation, especially around the aforementioned important boundary areas.…”

Automated polyp segmentation for colonoscopy images: A method based on convolutional neural networks and ensemble learning

“…Besides, it still has some applications in other research fields . SegNet also has many applications in the field of medical image segmentation . Moreover, its application in semantic segmentation is relatively mature .…”

“…27 SegNet also has many applications in the field of medical image segmentation. [28][29][30][31] Moreover, its application in semantic segmentation is relatively mature. [32][33][34][35] PSPNet is rarely used in medical image field.…”