Learning-based fully automated prediction of lumbar disc degeneration progression with specified clinical parameters and preliminary validation

Abstract

Background

Lumbar disc degeneration (LDD) may be related to aging, biomechanical and genetic factors. Despite the extensive work on understanding its etiology, there is currently no automated tool for accurate prediction of its progression.

Purpose

We aim to establish a novel deep learning-based pipeline to predict the progression of LDD-related findings using lumbar MRIs.

Materials and methods

We utilized our dataset with MRIs acquired from 1,343 individual participants (taken at the baseline and the 5-year follow-up timepoint), and progression assessments (the Schneiderman score, disc bulging, and Pfirrmann grading) that were labelled by spine specialists with over ten years clinical experience. Our new pipeline was realized by integrating the MRI-SegFlow and the Visual Geometry Group-Medium (VGG-M) for automated disc region detection and LDD progression prediction correspondingly. The LDD progression was quantified by comparing the Schneiderman score, disc bulging and Pfirrmann grading at the baseline and at follow-up. A fivefold cross-validation was conducted to assess the predictive performance of the new pipeline.

Results

Our pipeline achieved very good performances on the LDD progression prediction, with high progression prediction accuracy of the Schneiderman score (Accuracy: 90.2 ± 0.9%), disc bulging (Accuracy: 90.4% ± 1.1%), and Pfirrmann grading (Accuracy: 89.9% ± 2.1%).

Conclusion

This is the first attempt of using deep learning to predict LDD progression on a large dataset with 5-year follow-up. Requiring no human interference, our pipeline can potentially achieve similar predictive performances in new settings with minimal efforts.

Appendix: implantation process

Network architecture

The basic architecture of VGG-M [30] is adopted in our pipeline, which can be divided into two parts, the encoder and the classifier (Fig. 2). The encoder consists of a stack of alternately arranged convolutional layers and maxpooling layers [25], which is trained to extract the hierarchical feature from the input disc region. The classifier consists of two fully connected layers, which is trained to produce the pathology grade prediction based on the feature extracted by the encoder. The Rectified Linear Unit (ReLU) is served as the activation function of all convolutional and fully connected layers of the network, except the output, which is activated by the softmax function.

The input of our deep learning network is the disc region of a lumbar MRI, which is resized to $150\times 200\times 9$, where $9$ represents the middle $9$ slices of the MRI series. The output of the network is the probability prediction of each follow-up disc pathology grade, which is an $1\times 5$ array for the Pfirrmann grading and an $1\times 4$ array for the disc bulge or the Schneiderman score. The final grade prediction is the grade with the highest probability.

Since the original VGG-M is designed for the classification of natural images and has a large-scale dataset for the training process, two modifications are introduced in the network architecture for our specific task. First, the coordinate channels [34] are introduced in the input, which provide the relative location information and improve the position sensitivity of our network. Besides, the number of network parameters is reduced to accelerate the training process. More specifically, the channel numbers of the convolutional layers with the scale of $5\times 5$ and $3\times 3$ are reduced from 256 and 512 to 128 and 256, respectively, therefore the total number of network parameter is reduced from 6.5 to 2.8 M [30].

Training strategy

The basic idea of transfer learning [35] is adopted in the training process of the model, which is divided into two steps, pretraining and finetuning. In the pretraining step, the network is trained from the scratch for the pathology classification, which aims to enable the encoder of the network to extract the key features for the specific pathology from MRI. The label used in the pretraining step is the baseline disc pathology grade. Then, in the finetuning process, the network is further trained for pathology prediction, and the follow-up grade label is used to provide supervision. The parameters of the encoder in the pretrained network are preserved to inherit the ability of feature extraction, while the classifier is reinitialized for the new prediction task. The random on-the-fly data augmentation strategy is adopted in both training steps to reduce the risk of overfitting, which includes: (i) translation of $\pm 15\mathrm{\%}\times \mathrm{w}$ in x-axis and $\pm 10\mathrm{\%}\times \mathrm{w}$ in y-axis, where $\mathrm{w}$ represents the average width of the disc region (ii) rotation with $\pm 5^\circ$ (iii) rescaling with $1\pm 10\mathrm{\%}$ scaling factor.

Implementation details

13,130 disc samples are extracted from 2686 MRIs by the MRI-SegFlow for the validation of our method, and 6565 of them have the label of the follow-up pathology grades. The fivefold cross-validation strategy is employed. The samples with the follow-up grade labelled are equally divided into 5 subgroups. In each round of validation, one subgroup is selected as testing data and the other 4 subgroups combining with the samples without labels of follow-up grade are served as the training data. All training data are used in the pretraining step, while only the samples with the follow-up labels are used in the finetuning step. The testing data are invisible for the CNN model in the whole training process.

The mini-batch strategy is adopted in both two training steps with the batch size of 256 for pretraining and 64 for finetuning. The model reaches convergence in about 300 training epochs in the pretraining step and 100 in the finetuning step. In both steps, the learning rate is 0.001, Mean squared error is served as loss and the optimizer is Stochastic gradient descent. The data augmentation is applied in both training steps. TensorFlow 2.0 was used to implement the model with NVIDIA 2080Ti.