Figure 4

An example of heatmaps for an illustrative case that demonstrates a false positive in single-pose-based convolutional neural network (SP-CNN) models for a 78-year-old female patient without lumbar spinal stenosis (LSS). The 2 SP-CNN models (flexion and neutral posture) predicted the radiograph as LSS with probability on various spine levels and other irrelevant regions (red asterisks), but the multi-pose-based convolutional neural network (MP-CNN) model correctly predicted the radiographs as no LSS. This misprediction may have been caused by the high iliac crest line observed on both neutral and flexion images, which was comparable to that seen on extension images. MP-CNN model combines three postures and attenuates false positive lesions.