Figure 5 | Scientific Reports

Figure 5

From: Deep learning approach for automatic landmark detection and alignment analysis in whole-spine lateral radiographs

Figure 5

Learning speed of landmarks in different spinal areas. We categorised the landmarks of vertebral centres and femoral heads into four areas (cervical, thoracic, lumbar, and femoral heads) and plotted the per-area averaged standard deviation (SD) of the first-stage heatmaps against the training time (epoch), in log–log scale. From epoch 10 to epoch 100, we observed that the profiles of SDs had a tendency to decay as \({t}^{-\gamma }\). We fitted the decaying profiles in log–log scale (grey solid lines) and estimated \(\gamma =0.18\) (thoracic area); \(\gamma =0.58\) (lumbar area); \(\gamma =0.69\) (femoral heads); and \(\gamma =1.2\) (cervical area). For all fitted results, the adjusted \({R}^{2}>0.86\) and \(p\) value \(< {10}^{-4}\).

Back to article page