Figure 5
An example of heatmaps for an illustrative case that demonstrates extraction errors in single-pose-based convolutional neural network (SP-CNN) models for a 71-year-old female patient with lumbar spinal stenosis (LSS). The SP-CNN models in both neutral and flexion postures predicted LSS. However, the SP-CNN model in the neutral postures localized a probability of LSS on the sacrum (red asterisk) in addition to the foramen of L4/5. Additionally, the SP-CNN model in the flexion postures localized an even higher probability of LSS on the sacrum (blue asterisk) than on the foramen of L4/5. On the other hand, the multi-pose-based convolutional neural network (MP-CNN) model attenuated false positive and false negative results, such as iliac crest, bowel gas, spondylolisthesis, and disk height changes by postures, and accurately predicted LSS on L4/5.
