RETURN TO ABSTRACT LISTING


Objective:
MRI of the lumbosacral spine is performed to confirm clinical findings indicative of lumbar spinal stenosis (LSS). The reliability of radiological identification and grading could be enhanced by automation through the use of neural network models. Objective: To develop a neural network based imaging model to classify central canal stenosis and to demonstrate its efficacy as an accurate and consistent diagnostic tool.

Materials and Methods:
Measurements of anatomical objects and classifications of LSS were combined through two neural networks. The models were trained on 200 MRIs consisting of T-2 weighted images at each level. Evaluation of accuracy was performed on an external validation set of 116 MRI studies graded on a scale of absent, mild, moderate or severe by a panel of 7 radiologists. The standard was determined by majority and in case of disagreement, adjudicated by a further external radiologist. The interpretations were then compared to the interpretation of the model.

Results:
The model showed comparable performance to the average in terms of identifying stenosis as well as severity classification. For presence/absence, the sensitivity, specificity, AUC-ROC for the model was (0.88, 0.89, and 0.89), respectively, which was compared to the radiologist averages of (0.86, 0.86, 0.82). For multi-class severity, the model yielded metrics of (0.81, 0.90, and 0.85) which was compared to the radiologist averages of (0.77, 0.88, 0.82), respectively.

Conclusion:
The neural network model showed comparable performance to radiologists for the diagnosis of central canal stenosis. The integration of such models in the detection of LSS could bring higher accuracy, efficiency, consistency, and interpretability in diagnostic practices.