Skip to main content
. 2021 Aug 3;22(9):20–36. doi: 10.1002/acm2.13375

TABLE 3.

Summary of features used as inputs to ML/DL models for IMRT/VMAT QA outcome predictions.

Input features QA technique Feature selection technique Important features Reference
Plan complexity features 2D array detector LASSO 4 features: MU factor, aperture score, irregularity factor, and fraction of the plan delivered at the corners of a 40 × 40 cm2 field Valdes et al.33
2D array detector and EPID LASSO 7 features: irradiated area outline, Jaw position, fraction of the area receiving dose from penumbra, Duty cycle, irregularity factor, and others Valdes et al.34
Film dosimetry Dropout technique 4 features: MU values, the PTV volume, the rectum volume, and the overlapping region volume Tomori et al.35
3D array detector Pearson's correlation coefficient 28 features: plan complexity parameters (n = 18), machine type (n = 4), and photon beam energy (n = 6). Ono et al.46
3D array detector SVM recursive feature elimination 30 features: Linac output, MU factor, total number of control points, and others Granville et al.47
2D array detector Manually 54 features: MU value, union aperture area, plan area/irregularity/modulation, average leaf gap/dose rate/travel distance, modulation index of leaf speed/acceleration, and others Li et al.45
EPID Manually 10 features: modulation complexity score, beam irregularity, MUs/control point in a beam, maximum of xy jaw positions, edge metric, and others Lam et al.38
3D array detector Manually 54 features: plan modulation‐complexity and delivery‐characteristics Wang et al.51
2D array detector Extra‐trees, mutual information, and linear regression 100 features: aperture score, MU factor, edge metric, leaf gap/travel/motion, plan irregularity, plan modulation, and others Wall and Fontenot48
2D array detector Manually 6 features: leaf position, instantaneous velocity, movement away/toward the center, leaf movement status, the control point number, and the leaf bank Carlson et al.44
EPID Manually 14 features: leaf previous/current/next positions, dose fraction, gantry angle, leaf speed/acceleration, leaf gap, leaf movement status, and others Osman et al.41
EPID Manually 7 features: leaf velocity, acceleration, control point, dose rate, gravity vector, gantry velocity (VMAT), and gantry acceleration (VMAT) Chuang et al.43
Radiomics features EPID Wilcoxon rank‐sum Test 13 features: radiomics features (size zone metric) and intensity histogram metrics derived from gamma map images Wootton et al.37
2D array detector Dropout technique CNN features derived from fluence map images Interian et al.36
2D array detector Dropout technique CNN features derived from dose distribution images Tomori et al.50
EPID SVM recursive feature elimination 11 features: radiomics features e.g. contrast, uniformity, zone entropy, and others Ma et al.40
3D array detector Dropout technique CNN features derived from dose difference or gamma map images Kimura et al.52
EPID Random forest regression 11 features: radiomics intensity histogram features and the texture features derived from fluence difference map images Sakai et al 42
Combined features EPID Principal component analysis, and triplet networks 81 features: 17 texture features (hand‐crafted) from intensity histograms and size zone matrices) and 64 CNN features derived from gamma map images Nyflot et al.39
3D array detector Random forest 502 features: plan complexity features (e.g. aperture area/perimeter/irregularity, and others), radiomics features (shape, statistical information, and texture features), and clinical parameters (treatment site/machine, beam energy, and dose calculation algorithm) Hirashima et al.49

Abbreviations: CNN, convolutional neural network; EPID, electronic portal imaging device; LASSO, Least Absolute Shrinkage and Selection Operator; MU, monitor unit; PTV, planning target volume; SVM, support vector machine; VMAT, volumetric‐arc radiation therapy.