Table 4.
Statistical modeling and machine learning: features, models, and prediction outcomes
| Articles | Input features | Modeling methods | Prediction outcomes |
|---|---|---|---|
| Zhu et al.7, 9, 10, 13, 17, 20, 82 |
Volume features: PTV‐OAR overlap volume etc. Distance features: first three PCA components of Distance‐to‐Target Histogram |
Support vector regression Multivariate stepwise regression Model/regression tree |
The first three PCA components of DVH |
| Appenzoler et al.8 | OAR distance‐to‐PTV |
Sub‐DVH as basis functions of an OAR volume function of overlap subvolumes Function fitting using least squares minimization |
DVH |
| Lee et al.69, 70 | OVH values |
Logistic regression Linear regression K‐nearest neighbor |
Weight for an OAR constraint (Rectum, Bladder) |
| Yang et al.48 | Lx – distance from PTV that result in x% of overlap in OVH | Linear regression | Dx – dose received by x% of OAR volume |
| Fogliata et al. 11, 12, 15, 16, 18, 22, 23, 25, 26, 27, 28, 29, 31, 32, 34, 35, 37, 38, 39, 40, 41, 75 |
Volume features: PTV‐OAR overlap volume etc.Distance features: PCA components of Distance‐to‐target histogram Other unpublished features |
Multivariate regression (RapidPlan) | DVH |
| Nwankwo et al.59, 60 |
Distance‐to‐PTV Slice level |
Mean‐dose‐at‐distance function Mean dose standard deviation function Slice weight function |
Voxel dose |
| Amit et al.72 |
Beam‐independent features: tumor distribution, tumor height Beam‐dependent features: tumor‐organ overlap, beam distance, tumor projection shape |
Random forest regression | Beam angle |
| Liu et al.58 | 3D OAR structures |
Active shape model Active optical flow model |
Voxel dose |
| Wang et al.19 |
First two PCA component scores of OVH of OARs Z‐axis overlap index |
Stepwise multiple regression |
Mean lung dose Mean heart dose (forming a Pareto Front) |
| Yuan et al.73 | Beam number and angles | K‐medoids clustering | Standard beam bouquets |
| Cooper et al.50 | Distance to the tangent field edge | Logistic regression | Left anterior descending artery maximum dose |
| Kuo et al.51 |
Contralateral/ipsilateral lung volumes Ipsilateral normal/total lung volume MILD |
Linear regression |
Prescription dose MILD Prescription dose |
| Shiraishi et al.65 |
PTV volume Number of fields Azimuthal angle Elevation angle Distance from PTV Distance from OARs |
Artificial neural network (1 hidden layer with 10 nodes) | Voxel dose |
| Valdes et al.76, 77 | 78 complexity metrics: faction of MU per dose, jaw position, etc. | Poisson regression with Lasso regularization | Passing rate |
| Campbell et al.66 |
Geometric features: distance‐to‐PTV, distance to OARs, etc. Plan features: target volume, photon energy, etc. |
Artificial Neural Network (1 hidden layer with 25 nodes) | Voxel dose |
| Fan et al.30 |
Distance‐to‐PTV Angle with respect to origin of coordinate (center of CT) |
KDE | DVH |
| Powis et al.52 |
Fractional OAR‐PTV volume overlap Prescription dose |
Curve fitting | Mean rectum dose |
| Brown et al.78 |
Control point features Beam features Fraction group features Plan features |
Ensemble‐outlier filtering Normalized cut sampling SVM |
Classification (acceptable vs unacceptable plans) |
| Millunchick et al.54 | Fractional overlap of parotid with combined targets, and with 0.5 and 1.0 cm margins | Stepwise regression | Parotid mean dose |
PCA, principal component analysis; PTV, planning target volume; OAR, organ at risk; OVH, overlap volume histogram; MILD, mean ipsilateral lung dose; KDE, Kernel density estimate.