TABLE 1.
Summary of challenge submissions and performance metrics. All models achieved performance better than random (FDR < 5%)
| Rank | Description | AUROC | AP | C-index |
|---|---|---|---|---|
| 1 | Deep multitask logistic regression using EMR features and tumor volume. | 0.823 [0.777–0.866] | 0.505 [0.420–0.602] | 0.801 [0.757–0.842] |
| 2 | Fuzzy logistic regression (binary) and Cox proportional hazards model (risk prediction) using EMR features and tumor volume. | 0.816 [0.767–0.860] | 0.502 [0.418–0.598] | 0.746 [0.700–0.788] |
| 3 | Fuzzy logistic regression (binary) or Cox proportional hazards model (risk prediction) using EMR features and engineered radiomic features. | 0.808 [0.758–0.856] | 0.490 [0.406–0.583] | 0.748 [0.703–0.792] |
| 4 | Multitask logistic regression using EMR features. | 0.798 [0.748–0.845] | 0.429 [0.356–0.530] | 0.785 [0.740–0.827] |
| 5 | 3D convnet using cropped image patch around the tumor with EMR features concatenated before binary classification layer. | 0.786 [0.734–0.837] | 0.420 [0.347–0.525] | 0.774 [0.725–0.819] |
| 6 | 2D convnet using largest GTV image and contour slices with EMR features concatenated after additional nonlinear encoding before binary classification layer. | 0.783 [0.730–0.834] | 0.438 [0.360–0.540] | 0.773 [0.724–0.820] |
| 7 | 3D DenseNet using cropped image patch around the tumor with EMR features concatenated before multitask prediction layer. | 0.780 [0.733–0.824] | 0.353 [0.290–0.440] | 0.781 [0.740–0.819] |
| 8 | Multilayer perceptron (MLP) with SELU activation and binary output layer using EMR features. | 0.779 [0.721–0.832] | 0.415 [0.343–0.519] | 0.768 [0.714–0.817] |
| 9 | Two-stream 3D DenseNet with multitask prediction layer using cropped patch around the tumor and additional downsampled context patch. | 0.766 [0.718–0.811] | 0.311 [0.260–0.391] | 0.748 [0.703–0.790] |
| 10 | 2D convnet using largest GTV image and contour slices and binary output layer. | 0.735 [0.677–0.792] | 0.357 [0.289–0.455] | 0.722 [0.667–0.774] |
| 11 | 3D convnet using cropped image patch around the tumor and binary output layer. | 0.717 [0.661–0.770] | 0.268 [0.225–0.339] | 0.706 [0.653–0.756] |
| 12 | Fuzzy logistic regression (binary) and Cox proportional hazards model (risk prediction) using engineered radiomic features. | 0.716 [0.655–0.772] | 0.341 [0.272–0.433] | 0.695 [0.638–0.749] |