Table 3.

Case‐ and atlas‐based methods: similarity measure and knowledge transfer

Articles	Approach	Similarity measure	Knowledge transfer
Chanyavanich et al.56, 57, 61	Direct	Mutual information of the beam's eye view projections	Treatment parameters such as beam geometry and structure constraints and weights were transferred to the query case. The fluence maps were transferred after a deformable registration
Mishra et al.43 – Case‐based reasoning framework	Direct	Similarity is measured by both clinical variables such as clinical stage, Gleason score, and prostate‐specific antigen as well as rectum DVH values at five selected points	Dose constraints were transferred after adaptation
Petrovic et al.74	Direct	Further introduced knowledge‐light adaptation into the case retrieval process to improve case selection accuracy.	Dose constraints were transferred after adaptation
Wu et al.44, 45, 46, 47, 55	Direct	Based on the concept of OVH, which describes the fractional volume of an OAR that is within a specified distance from a PTV. For each OVH percent volume, the set of matching cases included all cases with smaller OVH values	The minimal DVH value at the percentage volume of the matching cases was transferred to the new case
Zhang et al.3 belongs to the automatic planning approach. We include it here because its beam selection is based on a database of prior clinical plans	Direct	Based on tumor location	Beam number and angles
Schreibmann et al.62, 71	Direct	Based on an iterative closest point registration algorithm and a score based on point to point distance	The beam settings and multileaf collimator positions for the best match were transferred to the new case
Zarepisheh et al.14	Direct	Based on machine learning algorithm that finds the best match of DVH curve using geometric features such as overlapping volume and mutual information
Zhou et al.49	Direct	The overlap area of OVH curves as the basis for similarity	Transferred DVH of OARs and PTV as optimization constraints.
Sheng et al.63 – Atlas‐based method	Direct	The generation of atlases and matching of a query case to the best atlas were both based on two specially designed features, the PTV and SV concaveness angle and the percent distance (from SV) to the PTV	Treatment parameters of the atlas case were transferred
Deshpande et al.24	Direct	Weighted sum of three difference values, the prescription dose differences, the OVH differences, and the difference of STS, which is a four‐dimensional histogram encoding the radial distance, azimuth, and elevation of PTV in relation to the center of an OAR. The difference of histograms is calculated by the earth mover's distance	The DVHs of top matching cases were presented for reviewing
McIntosh et al.64, 67, 68	Indirect	Each case in the database was associated with a contextual ARF that predicts dose at each voxel based on its location and image features. Each case was also associated with a random forest (pRF) that predicted the accuracy of the ARF for a new case based on its similarity to the new case's ARF. The set of matching cases had the smallest predicted errors from the associated pRF's.	The average predicted dose at voxel level from the ARF's of the matching cases was transferred as the voxel‐level dose of the new case
Li et al.33 – Atlas‐based method	Direct	A single atlas FDG‐PET volume was created from a set of prior clinical volumes using deformable registration of images and averaging of intensity values	The atlas was used as a template to generate a substructure of ABM within the pelvic bone marrow with a goal to improve sparing of ABM without manual contouring of ABM.
Valdes et al.53	Indirect	Differences between dosimetric indices of a database case and the predicted dosimetric indices of a query case must be smaller than predetermined thresholds. The predictions were based on boosted decision trees (random forest) that use features of anatomical information, medical records, treatment intent, and radiation transport.	Dosimetric information of matching cases was displayed

ABM, active bone marrow; ARF, atlas regression forest; PTV, planning target volume; OAR, Organ at Risk; OVH, overlap volume histogram; STS, spatial target signature; SV, seminal vesicle.