E.J.A. Verheijen, J. Chabros, C.L.A. Vleggeert-Lankamp
Dr. T. Smith
Lumbar spinal stenosis is a very common and debilitating medical condition caused by degenerative spinal changes due to aging. Patients experience severe leg pain, which significantly reduces physical functionality. Surgery can relieve symptoms; however, its efficacy remains unsatisfactory with a success rate of 69%.
In this study, a predictive model was developed using a Deep Learning approach to predict surgical success based on pre-operative MRI scans. A dataset containing MRI scans from 140 patients was used to train, validate, and test the model. Surgical success was determined at 26 weeks follow-up using a dichotomous outcome measure that comprised pain, disability, and perceived recovery. For every patient, axial MRI slices at the surgical lumbar level were extracted, cropped to 250×250 pixel images and augmented. A pre-trained ResNet50 model was utilized with a weighted binary cross entropy loss function. After testing, the model had an accuracy of 89.32% predicting surgical success correctly. The AUC-ROC was 94.73 indicating a low false positive rate. Using GradCAM heat maps, the model demonstrated to focus on the stenosed areas on the axial MRI scans. In conclusion, the model may outperform clinicians by approximately 20% in predicting surgical success for lumbar stenosis patients.
Lumbar spinal stenosis is a common back problem in the narrow tunnel in your spine where your nerves and spinal cord are. When this tunnel gets too tight, it puts pressure on your nerves and cord, causing pain and other issues. It is often seen in older people because as we age, our spines can wear down.
People with this problem often feel pain in their legs when they walk, as well as lower back pain, numbness, weakness, and tingling in their legs. Surgery can help, but it doesn’t always work well for everyone. About 7 out of 10 patients are satisfied after surgery.
We used a smart computer program to help decide who would benefit most from surgery. This computer program looked at MRI pictures of the spine. It learned from these pictures and also learned from how well people did after surgery. We trained and tested this computer program using MRI pictures from a total of 140 patients. The computer program was good at predicting who would do well after surgery, getting it right for about 9 out of 10 patients. This was better than what doctors could do, as the computer program was 20% more accurate.