Background
Methods
Training group
Study group
Image acquisition
Bone scan index
PET/CT index
-
Step 1: Convolutional neural network-based landmark detection
-
Step 2: Geometric model fittingPartly due to the limited training set, the convolutional neural network-based detectors produced a number of false positives but very few false negatives. To handle this, geometric models were used to prune false landmark detections and determine rough positions for the relevant anatomical structures. Essentially two types of models were used. The first was an iterative technique to track elongated bones such as ribs and clavicles. The second type was a classical active shape models used to find plausible positions for groups of landmarks.
-
Step 3: Convolutional neural network-based pixel-wise segmentationThe final step of the automated segmentation technique was the application of another convolutional neural network trained to perform pixel-wise segmentation of the CT image. The input to the network was not only the CT image but also a second channel with a rough segmentation based on an atlas registered using the aligned landmarks.
Statistical analyses
Results
Mean (SD) | Median (range) | Number of patients | |
---|---|---|---|
Age (years) | 73 (8.6) | 73 (53–92) | 48 |
PSA (μg/L) | 374 (874) | 84 (4–5740) | 48 |
Gleason score | 7.7 (1.5) | 8.0 (5–10) | 47 |
C-index | 95% CI |
p value | Hazard ratio | 95% CI |
p value | |
---|---|---|---|---|---|---|
BSI | 0.68 | 0.59–0.76 | <0.001 | 1.26 | 1.13–1.41 | <0.001 |
PET index | 0.69 | 0.60–0.78 | <0.001 | 1.17 | 1.06–1.29 | =0.002 |
PET15 index | 0.70 | 0.61–0.79 | <0.001 | 2.01 | 1.43–2.83 | <0.001 |