Table 5.
Comparison of the best ranked inference results across all stages of repair to expert segmentation.
| Input Config | FCN | Expert | |||||
|---|---|---|---|---|---|---|---|
| Images Only | Annulus | Commissures | Annulus + Commissures | Inter–User | Intra–User | ||
| Input Frames | CSP | CSP | Single-Phase | CSP | |||
| Resampling | No | Yes | Yes | Yes | N/A | N/A | |
| DSC | Anterior | 0.73[0.66–0.78] | 0.79[0.75–0.82] | 0.81[0.76–0.83] | 0.84[0.77–0.86] | 0.78[0.75–0.8] | 0.84[0.83–0.86] |
| Posterior | 0.64[0.57–0.71] | 0.75[0.73–0.8] | 0.77[0.71–0.8] | 0.8[0.74–0.83] | 0.73[0.69–0.76] | 0.77[0.71–0.8] | |
| Septal | 0.68[0.53–0.72] | 0.77[0.7–0.82] | 0.76[0.65–0.81] | 0.82[0.76–0.85] | 0.72[0.68–0.78] | 0.84[0.79–0.85] | |
| Merged | 0.76[0.71–0.8] | 0.86[0.81–0.88] | 0.82[0.75–0.84] | 0.85[0.82–0.88] | 0.81[0.79–0.84] | 0.87[0.85–0.87] | |
| Average | 0.67[0.61–0.72] | 0.77[0.73–0.81] | 0.78[0.68–0.81] | 0.81[0.75–0.84] | 0.74[0.7–0.78] | 0.82[0.79–0.83] | |
| MBD | Anterior | 0.96[0.53–1.26] | 0.63[0.48–0.89] | 0.48[0.34–0.67] | 0.36[0.27–0.51] | 0.83[0.52–1.05] | 0.49[0.38–0.54] |
| Posterior | 0.66[0.56–0.99] | 0.5[0.38–0.63] | 0.45[0.35–0.57] | 0.37[0.31–0.4] | 0.93[0.82–1.17] | 0.59[0.46–0.7] | |
| Septal | 0.73[0.59–0.87] | 0.46[0.36–0.73] | 0.51[0.37–0.67] | 0.37[0.26–0.48] | 0.92[0.64–0.97] | 0.62[0.41–1.07] | |
| Merged | 0.61[0.5–0.79] | 0.35[0.23–0.4] | 0.49[0.4–0.63] | 0.33[0.26–0.4] | 0.88[0.63–1.03] | 0.39[0.32–0.49] | |
| Average | 0.78[0.6–1.12] | 0.6[0.44–0.74] | 0.52[0.36–0.63] | 0.38[0.3–0.46] | 0.92[0.77–0.96] | 0.55[0.46–0.69] | |
DSC and MBD measurements are presented for individual leaflets (anterior, posterior, septal), merged leaflets and average across individual leaflets. All metric values are displayed as median [IQR]. The best ranked FCN input configuration across individual user input groups are displayed on the left and the expert segmenter results (inter-user and intra-user) are shown on the right. For individual leaflet averages, with providing more user input, the FCN results gradually improve. When providing the annulus contour, the FCN performs superior to the inter–user and is on par with the intra-user segmentation results. When providing the annulus contour and commissural landmarks combined, the FCN outperforms expert segmentation.