Table 1:
Performance comparison between classification and clustering with different encoders on sets of known disorders.
| Test set | Model | Images |
Supported syndromes | Null top-1 accuracy | Top-1 | Top-5 | Top-10 | Top-30 | |
|---|---|---|---|---|---|---|---|---|---|
| Gallery | Test | ||||||||
| F2G-frequent | Enc-F2G (softmax) | - | 2,669 | 299 | 0.33% | 35.94% | 52.45% | 63.91% | 78.13% |
| F2G-frequent | Enc-F2G | 19,950 | 2,669 | 299 | 0.33% | 21.06% | 39.62% | 49.12% | 67.98% |
| F2G-frequent | Enc-healthy | 19,950 | 2,669 | 299 | 0.33% | 10.69% | 23.69% | 31.46% | 50.80% |
|
| |||||||||
| F2G-rare | Enc-F2G | 2,348.8 | 1,183.3 | 816 | 0.12% | 13.66% | 23.62% | 29.56% | 40.94% |
| F2G-rare | Enc-healthy | 2,348.8 | 1,183.3 | 816 | 0.12% | 9.46% | 16.87% | 21.77% | 31.77% |
|
| |||||||||
| F2G-frequent | Enc-F2G | 22,298a | 2,669 | 1,115c | 0.09% | 20.15% | 37.81% | 46.85% | 64.21% |
| F2G-frequent | Enc-healthy | 22,298a | 2,669 | 1,115c | 0. 09% | 9.70% | 22.51% | 29.80% | 48.24% |
|
| |||||||||
| F2G-rare | Enc-F2G | 22,298.8b | 1,183.3 | 1,115c | 0. 09% | 7.07% | 14.19% | 17.67% | 24.41% |
| F2G-rare | Enc-healthy | 22,298.8b | 1,183.3 | 1,115c | 0. 09% | 4.02% | 8.84% | 11.73% | 16.61% |
The deep convolutional neural networks of Enc-F2G (softmax), Enc-F2G, and Enc-healthy have the same architecture. Training of Enc-F2G (softmax) and Enc-F2G was initiated with CASIA-WebFace and further fine-tuned on photos of patients in the Face2Gene frequent set. The Enc-F2G (softmax) model is the same as Enc-F2G, but using the softmax values of the layer instead of cosine distances between the FPDs in the CFPS. For the top-1 to top-30 columns, the best performance in each set is boldfaced. The numbers of images and syndromes in the rare set are averaged over ten splits. Enc-F2G outperformed Enc-healthy on both types of syndromes, showing the importance of fine-tuning on patient photos for learning facial dysmorphic features. The top-10 accuracy of Enc-F2G only drops by 2.27 percentage points (from 49.12% to 46.85%) after increasing the number of cases in the gallery and almost quadrupling the number of supported syndromes from 299 to 1,115.
Number of images in the frequent gallery + rare gallery.
Average of ten splits in the frequent gallery + rare gallery.
Number of syndromes in the frequent gallery + rare gallery.