Table 3 Methods for future contour and volumetric prediction in MRIgRT

Noorda et al., 2017 [45]

Combining subject-specific motion model with respiratory motion surrogate prediction for contour prediction (image prediction) in 4 healthy volunteers.

For prediction windows of 300 ms and 600 ms, the average DSC between the predicted liver contour and the ground truth were 0.944 ± 0.020 and 0.943 ± 0.021, respectively. Additionally, the vessel misalignment were 3.33 ± 0.90 mm and 3.53 ± 0.89 mm, respectively.

Ginn et al., 2020 [22]

Utilizing IR for contour prediction (image prediction) among 8 healthy volunteers and 13 cancer patients.

For prediction windows of 250 ms and 330 ms, the median distance between the predicted and the ground truth contour centroid positions using IR was 0.63 mm, while those of autoregressive linear prediction, linear extrapolation, and no-predictor were 0.84 mm, 1.15 mm, and 1.67 mm, respectively.

Romaguera et al., 2020 [55]

Utilizing ConvLSTM-STL for contour prediction (image prediction) among 12 healthy volunteers.

The median vessel misalignment were 0.45 mm and 0.57 mm for prediction windows of 320 ms and 640 ms, respectively.

Liu et al. 2021 [62]

Combining subject-specific motion model with PCA coefficients prediction for volumetric prediction among 8 patients with intrahepatic tumors.

For prediction window of 340 ms, median distances between predicted and ground truth centroid positions of targets were less than 1 mm on average.

Lombardo et al., 2023 [56]

This study compared LSTM-shift, ConvLSTM, ConvLSTM-STL for contour prediction using 88 patients from an internal cohort and the 3 patients from an external cohort.

For prediction window of 500 ms, the obtained DSC between the predicted target contour and the ground truth for LSTM-shift, ConvLSTM, ConvLSTM-STL and no-predictor were 0.92 ± 0.04, 0.91 ± 0.05, 0.91 ± 0.05, and 0.89 ± 0.05, respectively. Additionally, the RMSEs in the SI direction for these prediction methods were 1.3 ± 0.6 mm, 1.9 ± 1.0 mm, 1.9 ± 1.1 mm, and 2.8 ± 1.6 mm, respectively.

Romaguera et al., 2023 [48]

Combining population-based motion model with transformer network for contour and volumetric prediction among 25 healthy volunteers.

For image prediction at a prediction window of 450 ms, the obtained geometrical errors between the predicted image and the ground truth for ConvGRU, ConvLSTM, and transformer network were 1.60 ± 1.09 mm, 1.37 ± 0.92 mm, and 1.25 ± 0.74 mm, respectively.

For volumetric prediction at a prediction window of 450 ms, the obtained TREs between the predicted volumetric image and the ground truth for ConvGRU, ConvLSTM, and transformer network were 1.75 ± 1.19 mm, 1.66 ± 1.21 mm, and 1.56 ± 1.13 mm, respectively.