Deep Learning-based Volumetric Modulated Arc Therapy Dose Prediction with Uncertainty Quantification for Left- and Right-sided Breast Cancer: A Single-institution Feasibility Study

Abstract

Background:

Volumetric-modulated arc therapy (VMAT) enables precise tumor targeting while minimizing dose to organs at risk but requires time-consuming planning with quality variability. Most deep learning (DL) dose prediction studies focus on left-sided breast cancer, with limited exploration of right-sided cases or integrated uncertainty quantification. This study investigates the feasibility of a DL model for VMAT dose prediction in both left-and right-sided breast cancer, incorporating Monte Carlo dropout (MCD)-based uncertainty quantification within a single-institution dataset.

Methods:

A total of 68 clinically accepted breast cancer VMAT treatment plans (41 left-sided, 27 right-sided) were collected. A three-dimensional U-Net architecture was implemented. Two models with dropout rates of 0.25 (M0.25) and 0.50 (M0.50) were developed. MCD was applied for uncertainty quantification. Performance was evaluated using mean absolute error (MAE), gamma passing rate (GPR), and isodose Dice similarity coefficients.

Results:

The overall body MAE was 1.7 ± 0.11 Gy (M0.25) and 1.93 ± 0.11 Gy (M0.50). For the planning target volume (PTV) of 42.56 Gy, MAE was 0.93 ± 0.17 Gy (M0.25) and 0.88 ± 0.15 Gy (M0.50). The heart showed the lowest errors (M0.25: 0.87 ± 0.34 Gy; M0.50: 1.06 ± 0.38 Gy), while contralateral lung exhibited higher uncertainties. GPR at 3%/3 mm reached 85.5% for both models, with no significant difference between left- and right-sided cases. Uncertainty maps correlated with prediction errors (Pearson correlation =0.59) and identified regions with atypical anatomies.

Conclusions:

The proposed model supports the feasibility of VMAT dose prediction for both left-and right-sided breast cancer, while the uncertainty maps provide additional confidence information to assist clinical interpretation and decision-making. Uncertainty maps correlated with prediction error and flagged atypical cases. External validation with a larger, multi-institutional dataset is required before clinical implementation can be considered.

INTRODUCTION

Radiation therapy has been a cornerstone of cancer treatment since its introduction in the early 1990s.[1] Modern radiotherapy techniques now enable highly personalized plans with precisely conformal dose distributions, aimed at improving tumor control probability and minimizing normal tissue complication probability. Advanced inverse planning methods, such as volumetric-modulated arc therapy (VMAT), can deliver higher therapeutic doses to tumors while sparing surrounding healthy tissues, resulting in an improved therapeutic ratio.[2]

Breast cancer irradiation may particularly benefit from VMAT compared with traditional three-dimensional (3D) conformal radiotherapy because VMAT can improve planning target volume (PTV) coverage while minimizing high-dose exposure to organs at risk (OARs). These benefits are reflected in improved conformity and homogeneity indices.[3] However, the complexity of creating VMAT treatment plans increases workload, demands specialized expertise, and can limit patient capacity.[4] Traditional planning is also labor-intensive and prone to variability in plan quality,[5] becoming more so as radiation oncologists provide feedback and plans must be improved iteratively. Solutions such as knowledge-based planning (KBP) address some of these challenges using models trained on clinically accepted plans to generate estimated dose–volume histograms (DVHs) from contouring generated by radiation oncologists for a patient.[6,7,8] Although KBP enhances consistency and efficiency, it struggles with unique patient anatomies and cannot guarantee physically achievable dose distributions.[4,9,10]

In recent years, deep learning (DL) has emerged as a transformative tool in medical research, including image registration, segmentation, motion prediction, and toxicity prediction.[4] In radiotherapy, DL-based approaches have shown promise for dose distribution prediction, leveraging architectures such as U-Net to generate accurate, patient-specific dose predictions.[11] While applications in breast cancer VMAT remain limited, studies have demonstrated superior accuracy and efficiency compared with KBP.[4,12] However, most existing DL studies on dose prediction focus on left-sided breast cancer,[13,14,15] leaving right-sided cases comparatively underexplored despite presenting distinct dosimetric challenges.[16,17] Developing prediction models that address both lateralities within a single framework is therefore of practical interest for institutions treating unselected breast cancer populations.

Most DL-based dose prediction studies rely on deterministic models,[11] which cannot directly quantify prediction uncertainty for individual cases. Monte Carlo dropout (MCD)[18] provides a computationally efficient approximation to Bayesian neural networks by incorporating stochastic dropout layers during inference, thereby enabling uncertainty quantification, a capability that remains largely unexplored in radiotherapy dose prediction applications.[19] For medical tasks, it is important to know how confident the model is in its prediction and to decide whether or not to trust the model.[20] Incorporating uncertainty estimation can yield more robust and clinically relevant predictions.[18,20,21,22] This is particularly crucial in radiotherapy, where deterministic models may not identify cases with potentially unreliable predictions,[23] such as patients presenting with rare anatomical configurations or atypical target volume characteristics.

This study presents a preliminary, single-institution feasibility evaluation of a DL framework for VMAT dose prediction in breast cancer that integrates uncertainty quantification through MCD. The proposed model enables dose prediction for both left-and right-sided breast cancer while simultaneously providing uncertainty estimation. This method provides clinicians with essential confidence metrics to support clinical decision-making, while enhancing planning efficiency and maintaining consistently high treatment plan quality.

METHODS

Data collection

Data of breast cancer patients receiving radiation therapy using VMAT in Ramathibodi Hospital were retrospectively collected. The standard hypofractionated protocol delivers 42.56 Gy in 16 fractions, with a simultaneous integrated boost (SIB) of 49.60 Gy for high-risk PTVs. The treatment is conducted on two models of linear accelerators, TrueBeam and Clinac iX (Varian Medical Systems, Palo Alto, CA, USA), using 6 MV photon beams.

Planning computed tomography (CT) images were acquired using either the GE Optima 580W (GE Healthcare, Milwaukee, WI, USA) or Philips Big Bore (Philips Healthcare, Cleveland, OH, USA) CT simulators. Free-breathing CT images were reconstructed from helical acquisitions obtained at a tube voltage of 120 kVp, with an in-plane pixel size of 2.5 mm ×2.5 mm. Dose distributions were calculated in the Eclipse treatment planning system (TPS) (version 16.1, Varian Oncology Systems, Palo Alto, CA, USA) using the analytical anisotropic algorithm with a 2.5 mm calculation grid size.

For both left-and right-sided breast irradiation, photon beams were delivered through 2–3 partial arcs, with gantry angles spanning 130°–330° for left-sided targets and 40°–220° for right-sided targets. In selected cases requiring enhanced surface dose, a bolus is applied for 6–10 fractions. Although the partial-arc technique restricts modulation flexibility and may therefore reduce dose homogeneity, particularly in superficial regions, all plans are meticulously optimized to fulfill institutional objectives for target coverage and to comply with dose constraints for critical structures.[4,12]

In this study, a total of 68 clinically accepted breast cancer VMAT treatment plans from the Division of Radiation Oncology, Department of Diagnostic and Therapeutic Radiology, Faculty of Medicine Ramathibodi Hospital, Mahidol University, from January 2021 to September 2024 were collected for training, validation, and testing. Ethical approval for this study was obtained from the Institutional Review Boards (MURA2024/617). The dataset included 50 cases of whole breast/chest wall and regional nodes (axilla I–III, supraclavicular, internal mammary) treated with 42.5 Gy in 16 fractions, and 19 cases with an additional high-risk PTV boost to 49.60 Gy through a SIB, also delivered in 16 fractions. The plans were normalized such that 95% of the PTV received 95% of the prescription dose, following Ramathibodi’s in-house protocol. This protocol aligns with international recommendations[24,25] and prioritizes sparing of the heart, ipsilateral lung, and contralateral lung. Minor dose-priority variations were allowed based on individual clinical conditions and the overseeing radiation oncologist’s judgment. The detailed dataset for each patient is shown in Supplementary Table 1.

Supplementary Table 1.

Summary of patient data, detailing the irradiated side, computed tomography machine type, and the volumes for the planning target volume and various organs at risk

Case number	Irradiated side	Dataset	CT machine	ROI volume (cm3)
				PTV 42.56 Gy	PTV 49.6 Gy	Heart	Left lung	Right lung
1	Left	Train	GE	706		540	1079	1142
2	Left	Train	GE	1643	127	631	730	885
3	Left	Train	GE	821	55	994	997	1364
4	Left	Train	GE	770	155	458	792	933
5	Left	Train	GE	895	116	533	821	1078
6	Left	Train	GE	894		586	851	1386
7	Left	Train	GE	730		688	855	1243
8	Left	Train	GE	1306		604	713	891
9	Left	Train	GE	829		668	871	1256
10	Left	Train	GE	1039	74	601	771	1100
11	Left	Train	GE	1387		526	1262	1592
12	Left	Train	GE	839		544	793	1006
13	Left	Train	GE	889		516	1205	1407
14	Left	Train	GE	673		468	1026	1208
15	Left	Train	GE	1418	129	631	755	930
16	Left	Train	GE	989	28	707	1137	1238
17	Left	Train	GE	532		349	976	1142
18	Left	Train	GE	836		569	688	794
19	Left	Train	GE	802		513	996	1249
20	Left	Train	GE	873		557	796	1165
21	Left	Train	GE	722		353	1190	1355
22	Left	Train	GE	825		541	1060	1301
23	Left	Train	GE	2545		873	912	1308
24	Left	Train	GE	1281	53	670	846	1220
25	Left	Train	Phillip	861	3	535	733	1004
26	Left	Train	Phillip	517		393	1157	1340
27	Left	Train	Phillip	994		522	813	904
28	Left	Train	Phillip	898		426	789	988
29	Right	Train	GE	1092		633	885	959
30	Right	Train	GE	1045		631	639	849
31	Right	Train	GE	780		520	1068	1464
32	Right	Train	GE	801		483	918	1166
33	Right	Train	GE	595		492	1246	1514
34	Right	Train	GE	762		453	942	1332
35	Right	Train	GE	882	195	554	992	1283
36	Right	Train	GE	552		709	800	1152
37	Right	Train	GE	986		531	818	1309
38	Right	Train	GE	1119		534	785	1041
39	Right	Train	GE	1034		701	779	1091
40	Right	Train	GE	1307		665	1022	1239
41	Right	Train	Phillip	1119		520	765	980
42	Right	Train	Phillip	982		575	801	1118
43	Right	Train	Phillip	741	48	666	885	1095
44	Right	Train	Phillip	681		534	982	1259
45	Right	Train	Phillip	1056		553	946	1290
46	Left	Validation	Phillip	881		639	835	1159
47	Left	Validation	Phillip	1303		688	772	1182
48	Left	Validation	Phillip	500	39	531	1412	1732
49	Right	Validation	Phillip	743		615	1300	1703
50	Right	Validation	Phillip	693	54	562	1188	1252
L01	Left	Test	GE	658	11	526	849	1100
L02	Left	Test	GE	886	34	705	759	961
L03	Left	Test	GE	1132	103	441	1010	1094
L04	Left	Test	GE	681		563	966	1240
L05	Left	Test	GE	1175	73	602	1187	1561
L06	Left	Test	GE	1047		1006	714	1100
L07	Left	Test	GE	621		542	739	882
L08	Left	Test	Phillip	701		617	1366	1564
L09	Left	Test	Phillip	904		516	940	1171
L10	Left	Test	Phillip	1037	8	743	533	866
R01	Right	Test	GE	730		420	817	1099
R02	Right	Test	Phillip	854		607	595	982
R03	Right	Test	Phillip	845		505	1544	1838
R04	Right	Test	Phillip	882	4	658	1139	1257
R05	Right	Test	Phillip	632		520	792	1269
R06	Right	Test	Phillip	1120		682	729	1167
R07	Right	Test	Phillip	1172		749	905	1277
R08	Right	Test	Phillip	894		587	766	911

A blank entry for PTV 49.6 Gy indicates the absence of a high-risk PTV. PTV: Planning target volume, CT: Computed tomography, ROI: Regions of interest

Pre-processing

Treatment planning data, including planning CT images, structure sets, and dose distributions, were exported from the TPS and preprocessed using DicomRTTool[26] and SimpleITK[27] on Python (version 3.10). All CTs and their corresponding DICOM-RT files were down-sampled to a 128 × 128 × 128 matrix with 3.5 mm ×3.5 mm pixel dimensions and a slice thickness of 3 mm. In cases involving a partial bolus insertion fraction, the total cumulative dose was calculated and used as the final clinical dose distribution.

The Hounsfield Unit of the planning CT, ranging from −1024 to 3200, was mapped to (0, 1). Non-anatomical areas outside the body contour were set to zero to eliminate irrelevant information. Binary contour masks were generated for the PTV 42.56 Gy, PTV 49.6 Gy, heart, left lung, right lung, and body. The intersection between PTV 42.56 Gy and PTV 49.6 Gy was removed from the PTV 42.56 Gy binary, and all regions of interest (ROIs) were cropped out of the body contour. Dose distributions were normalized by dividing by 42.56 (prescription dose).

Model architecture and training

A 3D U-Net architecture[28] was employed to predict the clinically accepted dose distribution. The input layer comprised seven channels: the planning CT, two PTV masks (42.56 Gy and 49.60 Gy), and four OAR masks (heart, left lung, right lung, and body). The network utilized an encoder and decoder design with six levels, each containing two 3 × 3 × 3 convolutional layers. Down-sampling was accomplished through 2 × 2 × 2 max pooling, while 2 × 2 × 2 up-sampling layers in the decoder path restored spatial resolution [Supplementary Figure 1 ^{(67.9KB, tif)} ]. Skip connections from corresponding encoder layers preserved spatial detail crucial for accurate dose prediction.

Batch normalization followed each convolutional layer, and dropout layers were inserted after both max pooling and up-sampling operations to mitigate overfitting and provide uncertainty estimates. ReLU[29] was employed as the activation function following each convolution, introducing non-linearity essential for capturing complex patterns. The mean squared error (MSE) loss function was minimized using the Adam optimizer.[30]

The dataset was randomly partitioned into 45 plans (28 left, 17 right) for training, 5 plans (3 left, 2 right) for validation, and 18 plans (10 left: L01–L10; 8 right: R01–R08) for testing. Two identical models with different dropout rates of 0.25 (M0.25) and 0.50 (M0.5) were trained. Each model was trained for a maximum of 500 epochs on an NVIDIA GeForce RT×4090 GPU.

Evaluation

During the inference phase, MCD was activated to generate 35 stochastic predictions per test case, which were then averaged to yield a robust final dose distribution.

Mean absolute error (MAE), MSE, dosimetric indices (DIs), gamma passing rate (GPR), and isodose dice similarity coefficients (iDSCs) were used as performance metrics. The GPR analysis was conducted using 3%/3 mm at the 80% prescription dose threshold, consistent with prior breast dose prediction studies,[12] along with more stringent (1%/1 mm, 2%/2 mm) criteria. A high GPR indicates concordance between the clinical and predicted dose distributions. iDSCs were calculated to assess spatial overlap between the clinical and predicted dose distributions at various isodose levels. For an isodose level, the iDSC is defined as are the isodose volumes of D Gy from the clinical plan and model-predicted dose distribution, respectively.

Uncertainty quantification and evaluation

The uncertainty associated with model predictions was quantified using the voxel-wise standard deviation calculated from 35 stochastic predictions. Due to the property of MCD, raw uncertainty estimates provide only relative measures, as their magnitudes are confounded by parameters such as dropout rate.[31] To improve interpretability, raw uncertainty values were rescaled following the approach of Nguyen et al.[31] Specifically, this study applied the following calculation on the validation data to determine a scaling factor for a given ROI with voxels in the ROI. This approach ensures that the uncertainty values provide a consistent and standardized measure across different regions, allowing for meaningful comparisons between ROIs. To ensure that uncertainty estimates reflect physical prediction error, a scaling factor was computed on the validation set. At the voxel level, the ideal scaling parameter m_i represents the ratio between the absolute dose error and the predicted standard deviation:

where D_predicti and D_plani are the dose from the plan and model prediction at the voxel i, respectively. σ_i is the standard deviation of voxel i. Because m_i varies spatially, individual voxel ratios within a given ROI were aggregated into a regional scaling factor, m_σ,ROI.

By scaling m_σ,ROI, the model’s relative variance is rescaled to match the expected magnitude of absolute dose error within each anatomical structure. These ROI-specific factors were computed on the validation set and subsequently used for test-set evaluation.

The uncertainty maps were used to estimate dose distribution bounds by:

where D_i is the final radiation dose distribution predicted by the model at location i, and D_Li and D_Hi are the higher and lower bounds of the achievable dose distribution, respectively. To evaluate the reliability of these uncertainty intervals, coverage probability (P(y))[32] was defined as the proportion of dose points falling within these uncertainty intervals, calculated as the ratio of voxels within each ROI. The values of P(y)provide insight into the coverage of the predicted dose and its uncertainty interval by identifying the proportion of dose points within the model’s uncertainty bounds. A higher coverage probability indicates better agreement between the predicted and actual dose distributions. Finally, DVH for each contour structure was calculated.

RESULTS

Prediction accuracy

Figure 1 presents a comparison between D_plan and D_predict for L04. The figure also includes the corresponding percentage dose difference maps. For test case L04, both models effectively replicate D_plan across high, medium, and low dose levels. The difference maps reveal localized discrepancies, predominantly within the PTV boundary, particularly within the ipsilateral lung region where doses were underestimated by more than 20%. The differences observed in M0.25 and M0.50 indicate comparable performance between the two models.

Figure 1.

Comparison of D_plan and D_predict absolute dose distributions for test case L04 from model M0.25 (left) and M0.50 (right). In the difference map, gray indicates the planning target volume structure

MAE and MSE are shown in Table 1. The errors for PTV 42.56 Gy are lower compared to the OARs. However, for PTV 49.60 Gy, higher errors were observed. With the limited number of test cases with the high-risk PTV, the average MAE and MSE values are reported for the total average combining left- and right-sided breast irradiation. Among the OARs analyzed, the heart demonstrated the most accurate dose predictions with the lowest MAE values of 0.87 ± 0.34 Gy for the M0.25 model (Lt: 0.84 ± 0.27 Gy, Rt: 0.91 ± 0.4 Gy) and 1.06 ± 0.38 Gy for the M0.50 model (Lt: 1.04 ± 0.31 Gy, Rt: 1.08 ± 0.45 Gy). In contrast, the contralateral lung exhibited the largest prediction uncertainties among all OARs evaluated, with substantially higher MAE values of 1.81 ± 0.49 Gy and 2.13 ± 0.64 Gy for the M0.25 and M0.50 models, respectively. In addition, to assess evaluation stability across the test cohort, 95% confidence intervals derived from bootstrap resampling (1,000 iterations) for MAE, MSE, and 3D GPRs are provided in Supplementary Table 2.

Table 1.

Average of mean absolute error and mean squared error calculated for planning target volume 42.56 Gy, planning target volume 49.60 Gy, heart, ipsilateral lung, contralateral lung, and body of M0.50 and M0.25 (Left: Left-sided breast irradiation; Right: Right-sided breast irradiation)

ROI	M0.25			M0.50
	Left	Right	Total	Left	Right	Total
MAE (Gy)
PTV 42.56 Gy	0.92±0.13	0.95±0.2	0.93±0.17	0.89±0.15	0.87±0.14	0.88±0.15
PTV49.60 Gy	4.68±1.74	1.02±0.0	4.07±2.09	4.88±2.02	0.58±0.0	4.16±2.45
Heart	0.84±0.27	0.91±0.4	0.87±0.34	1.04±0.31	1.08±0.45	1.06±0.38
Ipsilateral lung	1.56±0.63	1.92±0.35	1.72±0.56	1.67±0.66	2.36±0.35	1.97±0.64
Contralateral lung	2.09±0.32	1.46±0.44	1.81±0.49	2.58±0.36	1.56±0.45	2.13±0.64
Body	1.71±0.07	1.68±0.14	1.7±0.11	1.96±0.06	1.9±0.14	1.93±0.11
MSE (Gy2)
PTV 42.56 Gy	1.47±0.35	1.56±0.58	1.51±0.47	1.35±0.37	1.39±0.44	1.36±0.4
PTV49.60 Gy	27.14±18.1	1.48±0.0	22.87±19.09	30.51±24.51	0.5±0.0	25.51±25.01
Heart	1.87±1.34	2.51±2.36	2.16±1.89	2.7±1.79	3.22±2.81	2.93±2.31
Ipsilateral lung	6.3±4.51	7.27±2.08	6.73±3.67	6.92±5.0	10.86±2.48	8.67±4.52
Contralateral lung	8.13±2.15	5.4±2.63	6.92±2.73	12.24±2.82	5.82±2.72	9.39±4.23
Body	7.77±0.64	7.68±1.16	7.73±0.91	9.61±0.63	9.27±1.33	9.46±1.02

PTV: Planning target volume, MAE: Mean absolute error, MSE: Mean squared error, ROI: Regions of interest

Supplementary Table 2.

Model dose prediction performance on the independent test set for M0.25 and M0.50

ROI	Metric	M0.25 mean	M0.25 95% CI	M0.50 Mean	M0.50 95% CI
PTV 42.56 Gy	MAE (Gy)	0.93	0.85–1.01	0.88	0.81–0.94
	MSE (Gy2)	1.51	1.31–1.73	1.36	1.19–1.55
PTV 49.60 Gy	MAE (Gy)	4.07	2.65–5.66	4.16	2.45–6.06
	MSE (Gy2)	22.87	9.60–36.33	25.51	9.86–45.45
Heart	MAE (Gy)	0.87	0.72–1.03	1.06	0.89–1.24
	MSE (Gy2)	2.16	1.39–3.10	2.93	1.97–4.08
Ipsilateral lung	MAE (Gy)	1.72	1.46–1.97	1.97	1.67–2.27
	MSE (Gy2)	6.73	5.12–8.49	8.67	6.68–10.84
Contralateral lung	MAE (Gy)	1.81	1.58–2.04	2.13	1.83–2.43
Contralateral lung	MSE (Gy2)	6.92	5.08–8.87	9.39	6.49–12.51
Body	MAE (Gy)	1.7	1.62–1.77	1.93	1.85–2.00
	MSE (Gy2)	7.73	7.11–8.35	9.46	8.76–10.15
Gamma passing rate (80% Cutoff)	GPR 1%/1 mm (%)	43.81	40.42–47.01	44.13	40.85–47.21
	GPR 2%/2 mm (%)	69.93	66.52–73.08	70.48	66.86–73.91
	GPR 3%/3 mm (%)	85.54	82.90–87.97	85.54	82.68–88.24

Values are reported as the mean with 95% CI estimated through bootstrapping (1000 resamples). CI: Confidence interval, PTV: Planning target volume, MAE: Mean absolute error, GPR: Gamma passing rate, MSE: Mean squared error, ROI: Regions of interest

Table 2 presents the 3D GPRs for M0.25 and M0.50 models across 18 test cases evaluated using 1%/1 mm, 2%/2 mm, 3%/3 mm, and 5%/5 mm criteria with an 80% dose cutoff. Both models demonstrated improved performance with relaxed tolerance criteria. Under the stringent 1%/1 mm criterion, average GPRs were 43.81 ± 7.33% and 44.13 ± 7.06% for M0.25 and M0.50, respectively. Performance substantially improved at the 2%/2 mm criterion (M0.25: 69.93 ± 7.50%, M0.50: 70.48 ± 7.77%) and further increased at 3%/3 mm (both models: 85.54%, with standard deviations of 5.60% and 6.23% for M0.25 and M0.50, respectively).

Table 2.

Average percentage three-dimensional gamma passing rates of M0.25 and M0.50 using 1%/1 mm, 2%/2 mm, 3%/3 mm, and 5%/5 mm criteria with 80% dose cutoff

	1%/1 mm (%)		2%/2 mm (%)		3%/3 mm (%)
	M0.25	M0.50	M0.25	M0.50	M0.25	M0.50
Left	44.80±7.72	45.06±8.56	71.35±6.70	71.30±7.65	86.47±4.66	86.26±5.60
Right	42.58±7.13	42.97±4.90	68.16±8.52	69.46±8.32	84.38±6.74	84.64±7.22
Total average	43.81±7.33	44.13±7.06	69.93±7.50	70.48±7.77	85.54±5.60	85.54±6.23
P	0.7	0.53	0.41	0.63	0.46	0.76

Analysis of laterality effects revealed marginally higher GPRs for left-sided compared to right-sided breast treatments across all criteria, with differences of approximately 2% at the 3%/3 mm level. However, Mann–Whitney U tests demonstrated these differences were not statistically significant (all P > 0.05). Similarly, Wilcoxon signed-rank tests comparing M0.25 and M0.50 performance yielded non-significant results across all criteria (1%/1 mm: P =0.67, 2%/2 mm: P =0.37, 3%/3 mm: P =0.52), indicating equivalent overall performance between the two dropout configurations.

The analysis of DIs between predicted and planned doses reveals general agreement between models M0.25 and M0.50 and clinical plans, though some notable discrepancies exist [Supplementary Table 3]. Statistical analysis was performed using the Wilcoxon signed-rank test. For the PTV, D_predict from M0.50 demonstrated concordance with D_plan across all metrics for both left-and right-sided breast irradiation, with no statistically significant difference observed (P > 0.05), except for D_95%. This exception is expected due to the low variance in planned doses resulting from normalization. In contrast, M0.25 showed significant deviations in left-side for D95% and D98% (P = 0.02 and 0.037, respectively) and right-side for D50% (P = 0.016). However, the magnitude of these differences was minimal.

Supplementary Table 3.

Comparison of dosimetric indices (mean ± standard deviation) between the plan and predicted dose in planning target volume and key regions of interest (heart, ipsilateral lung, and contralateral lung) for M0.25 and M0.50.

Model	ROIs	DI	Left-sided breast irradiation			Right-sided breast irradiation
			Plan mean	Predicted mean	P	Plan mean	Predicted mean	P
M0.25	PTV	D2% (Gy)	47.29±2.69	47.7±3.83	0.557	44.93±0.34	45.05±0.57	0.844
		D50% (Gy)	43.59±0.48	43.75±0.2	0.275	43.29±0.22	43.79±0.31	0.016
		D95% (Gy)	40.36±0.02	40.82±0.43	0.02	40.37±0.01	40.44±0.55	0.844
		D98% (Gy)	39.14±0.26	39.67±0.46	0.037	39.2±0.2	39.18±0.76	0.945
		D99% (Gy)	38.3±0.49	38.84±0.46	0.105	38.33±0.34	38.3±0.83	1
		Dmean (Gy)	43.43±0.49	43.6±0.39	0.16	43.02±0.19	43.33±0.27	0.055
		Dmax (Gy)	49.37±3.15	50.16±4.06	0.232	46.88±1.88	46.94±1.85	0.945
	Heart	V17Gy (%)	6.8±2.82	6.27±3.01	0.77	0.11±0.17	0.46±0.47	0.125
		V35Gy (%)	0.43±0.41	0.51±0.51	0.492	0.0±0.0	0.0±0.0	1
		D15% (Gy)	10.83±2.28	10.16±2.48	0.492	5.67±0.69	6.13±1.17	0.383
		D20% (Gy)	9.09±2.01	8.56±2.01	0.557	5.09±0.62	5.34±1.02	0.461
		Dmean (Gy)	6.92±1.14	6.69±1.23	0.695	3.86±0.45	3.85±0.71	0.945
	Ipsilateral lung	V17Gy (%)	30.15±4.35	33.51±6.32	0.064	29.6±5.13	30.95±5.42	0.641
		D15% (Gy)	30.14±3.12	31.15±3.96	0.16	28.61±3.49	29.45±3.17	0.547
		D20% (Gy)	25.54±3.5	27.06±4.24	0.064	23.97±3.46	24.88±3.46	0.641
		D35% (Gy)	13.99±2.58	16.11±3.35	0.064	14.17±2.68	14.91±2.67	0.641
		D50% (Gy)	8.0±1.17	9.7±2.44	0.037	9.08±1.69	9.48±1.86	0.547
		Dmean (Gy)	13.56±1.32	14.57±1.96	0.027	13.78±1.64	14.08±1.6	0.461
	Contra-lateral lung	D20% (Gy)	4.96±1.5	4.45±0.52	0.625	4.94±1.25	3.88±0.31	0.039
		D35% (Gy)	3.82±1.2	3.25±0.31	0.193	3.56±0.92	2.98±0.14	0.25
		D50% (Gy)	3.06±0.89	2.65±0.22	0.193	2.7±0.65	2.39±0.19	0.383
		Dmean (Gy)	3.61±0.88	3.25±0.31	0.275	3.39±0.67	2.82±0.36	0.039
M0.50	PTV	D2% (Gy)	47.29±2.69	46.78±3.06	0.275	44.93±0.34	44.58±0.39	0.109
		D50% (Gy)	43.59±0.48	43.56±0.21	0.846	43.29±0.22	43.46±0.28	0.25
		D95% (Gy)	40.36±0.02	40.77±0.3	0.01	40.37±0.01	40.08±0.68	0.312
		D98% (Gy)	39.14±0.26	39.55±0.48	0.084	39.2±0.2	38.77±0.87	0.312
		D99% (Gy)	38.3±0.49	38.65±0.66	0.322	38.33±0.34	37.84±0.98	0.25
		Dmean (Gy)	43.43±0.49	43.38±0.4	0.922	43.02±0.19	42.99±0.28	0.641
		Dmax (Gy)	49.37±3.15	49.08±3.32	0.492	46.88±1.88	46.14±1.29	0.148
	Heart	V17Gy (%)	6.8±2.82	5.81±3.41	0.432	0.11±0.17	0.39±0.41	0.125
		V35Gy (%)	0.43±0.41	0.53±0.53	0.695	0.0±0.0	0.0±0.0	1
		D15% (Gy)	10.83±2.28	9.94±2.44	0.322	5.67±0.69	6.21±1.24	0.312
		D20% (Gy)	9.09±2.01	8.29±1.74	0.322	5.09±0.62	5.45±1.11	0.383
		Dmean (Gy)	6.92±1.14	6.61±1.21	0.557	3.86±0.45	3.87±0.76	1
	Ipsilateral lung	V17Gy (%)	30.15±4.35	30.78±5.05	0.695	29.6±5.13	29.88±5.52	0.844
		D15% (Gy)	30.14±3.12	30.27±3.87	1	28.61±3.49	29.01±3.41	0.641
		D20% (Gy)	25.54±3.5	25.75±3.93	1	23.97±3.46	24.31±3.69	0.641
		D35% (Gy)	13.99±2.58	14.54±2.87	0.695	14.17±2.68	14.38±2.68	0.844
		D50% (Gy)	8.0±1.17	8.75±1.82	0.16	9.08±1.69	9.18±1.75	0.547
		Dmean (Gy)	13.56±1.32	13.99±1.76	0.322	13.78±1.64	13.75±1.48	0.641
	Contra-lateral lung	D20% (Gy)	4.96±1.5	4.54±0.58	0.77	4.94±1.25	3.81±0.51	0.039
		D35% (Gy)	3.82±1.2	3.54±0.41	0.922	3.56±0.92	2.95±0.31	0.148
		D50% (Gy)	3.06±0.89	2.92±0.31	0.77	2.7±0.65	2.34±0.31	0.25
		Dmean (Gy)	3.61±0.88	3.36±0.4	0.625	3.39±0.67	2.73±0.37	0.016
M0.25	PTV	D2% (Gy)	47.38±2.71	47.27±4.04	0.695	44.99±0.34	45.05±0.95	0.945
		D50% (Gy)	43.68±0.49	43.34±0.43	0.275	43.36±0.23	43.79±0.75	0.109
		D95% (Gy)	40.43±0.0	40.43±0.0	1	40.43±0.0	40.43±0.0	1
		D98% (Gy)	39.22±0.26	39.3±0.1	0.492	39.26±0.19	39.17±0.26	0.547
		D99% (Gy)	38.38±0.48	38.48±0.2	0.695	38.39±0.34	38.29±0.35	0.547
		Dmean (Gy)	43.51±0.51	43.2±0.64	0.322	43.09±0.19	43.33±0.63	0.25
		Dmax (Gy)	49.47±3.17	49.71±4.18	0.557	46.95±1.88	46.96±2.38	0.945
	Heart	V17Gy (%)	6.82±2.83	6.14±2.93	0.625	0.11±0.17	0.48±0.49	0.125
		V35Gy (%)	0.44±0.41	0.46±0.47	0.82	0.0±0.0	0.0±0.0	1
		D15% (Gy)	10.85±2.28	10.05±2.41	0.375	5.68±0.7	6.13±1.17	0.383
		D20% (Gy)	9.11±2.01	8.48±1.96	0.432	5.09±0.62	5.34±1.02	0.461
		Dmean (Gy)	6.93±1.14	6.62±1.19	0.557	3.86±0.45	3.86±0.71	1
	Ipsilateral lung	V17Gy (%)	30.19±4.35	33.26±6.21	0.084	29.64±5.14	30.94±5.48	0.641
		D15% (Gy)	488.39±76.19	487.64±70.35	0.625	379.32±47.75	415.41±59.22	0.148
		D20% (Gy)	414.08±75.83	423.39±72.79	0.846	317.05±43.24	350.1±52.58	0.148
		D35% (Gy)	226.93±50.51	253.21±64.05	0.625	186.65±30.88	209.42±35.28	0.195
		D50% (Gy)	128.91±20.21	153.08±48.02	0.375	119.02±12.91	132.98±23.25	0.25
		Dmean (Gy)	13.58±1.33	14.43±1.9	0.064	13.81±1.64	14.08±1.64	0.641
	Contra-lateral lung	D20% (Gy)	157.81±40.1	140.56±21.96	0.322	171.2±52.12	151.82±31.1	0.312
		D35% (Gy)	121.29±31.73	102.4±12.46	0.193	123.39±38.2	116.97±25.53	0.844
		D50% (Gy)	97.16±23.13	83.49±8.72	0.131	93.51±27.54	93.36±20.16	0.844
		Dmean (Gy)	3.62±0.88	3.21±0.29	0.275	3.39±0.67	2.82±0.36	0.023
M0.50	PTV	D2% (Gy)	47.38±2.71	46.39±2.89	0.037	44.99±0.34	44.98±0.89	0.742
		D50% (Gy)	43.68±0.49	43.2±0.29	0.049	43.36±0.23	43.85±0.73	0.109
		D95% (Gy)	40.43±0.0	40.43±0.0	1	40.43±0.0	40.43±0.0	1
		D98% (Gy)	39.22±0.26	39.22±0.22	0.77	39.26±0.19	39.11±0.25	0.25
		D99% (Gy)	38.38±0.48	38.33±0.43	1	38.39±0.34	38.16±0.41	0.25
		Dmean (Gy)	43.51±0.51	43.02±0.33	0.049	43.09±0.19	43.38±0.62	0.195
		Dmax (Gy)	49.47±3.17	48.66±3.12	0.16	46.95±1.88	46.56±1.81	0.312
	Heart	V17Gy (%)	6.82±2.83	5.7±3.33	0.322	0.11±0.17	0.41±0.44	0.125
		V35Gy (%)	0.44±0.41	0.47±0.48	0.432	0.0±0.0	0.0±0.0	1
		D15% (Gy)	10.85±2.28	9.84±2.38	0.275	5.68±0.7	6.21±1.26	0.312
		D20% (Gy)	9.11±2.01	8.21±1.7	0.275	5.09±0.62	5.45±1.12	0.312
		Dmean (Gy)	6.93±1.14	6.54±1.18	0.492	3.86±0.45	3.88±0.77	0.945
	Ipsilateral lung	V17Gy (%)	30.19±4.35	30.53±4.95	0.922	29.64±5.14	29.89±5.59	0.844
		D15% (Gy)	488.39±76.19	474.36±72.24	0.322	379.32±47.75	409.93±65.91	0.148
		D20% (Gy)	414.08±75.83	403.63±73.4	0.322	317.05±43.24	342.47±57.43	0.383
		D35% (Gy)	226.93±50.51	229.76±61.02	0.77	186.65±30.88	201.3±32.77	0.25
		D50% (Gy)	128.91±20.21	138.57±39.2	0.846	119.02±12.91	128.41±20.24	0.383
		Dmean (Gy)	13.58±1.33	13.85±1.71	0.557	13.81±1.64	13.76±1.52	0.742
	Contra-lateral lung	D20% (Gy)	157.81±40.1	142.79±20.05	0.625	171.2±52.12	148.89±34.44	0.312
		D35% (Gy)	121.29±31.73	111.56±14.05	0.625	123.39±38.2	115.77±27.64	0.742
		D50% (Gy)	97.16±23.13	91.93±10.42	0.922	93.51±27.54	91.16±21.1	1
		Dmean (Gy)	3.62±0.88	3.32±0.39	0.557	3.39±0.67	2.73±0.37	0.016

PTV: Planning target volume, ROI: Regions of interest, DI: Dosimetric indices

Heart dosimetry remained well-preserved across both models, with no statistically significant differences observed for any cardiac metrics in either laterality. Mean heart doses were consistently predicted within clinically acceptable ranges (left-sided: plan 6.92 ± 1.14 Gy vs. 6.69 ± 1.23 Gy for M0.25 and 6.61 ± 1.21 Gy for M0.50), and the critical V35 Gy metric showed minimal variation. For the ipsilateral lung, both models slightly overestimated dose. Significant differences were observed for the left side in D50% and D_mean with M0.25, whereas M0.50 showed no statistically significant differences for any ipsilateral lung metric. The contralateral lung showed the prediction challenges, particularly for right-sided irradiation where both models significantly underestimated D20% and D_mean (P < 0.05 for both M0.25 and M0.50).

Figure 2 shows the spatial agreement between D_plan and D_predict in terms of iDSC for all test cases and a high-risk PTV of 49.60 Gy for model M0.25. For most cases, the iDSC is above 0.8 except for case R05, which has an iDSC below 0.8 around lower isodose levels. The average iDSC of all cases starts at 0.88, crosses 0.90 at about 55%, and peaks above 0.93 at 90%–95% isodose. For high-risk PTV cases, the iDSC shows higher deviations around the high-risk PTV. Model M0.5 also shows similar trends.

Figure 2.

The isodose dice similarity coefficient (iDSC) as a function of the % isodose level of model M0.25. The plot on the left displays the iDSC for all test cases (L01–L10, R01–R08). The plot on the right shows the iDSC for the subset of test cases specifically treated to a prescription dose of 49.60 Gy (L01, L02, L03, L04, L05, L06, L07). Dashed lines show individual cases, and the red solid line is the mean iDSC

Uncertainty quantifications

Figure 3 presents different maps of D_plan and D_predict, along with the corresponding voxel-wise raw uncertainty maps. In case R05, unfamiliar anatomical structures (arm) lead to elevated uncertainty. This demonstrates the ability of MCD to highlight regions of low confidence when atypical anatomy is encountered. Across all cases, prediction errors were visually correlated with the corresponding uncertainty maps. A voxel-wise Pearson correlation test confirmed this relationship, yielding average correlations between the absolute difference map and uncertainty of 0.59 for M0.25 and M0.50 (P < 0.05). In addition, voxel-wise standard deviations across 35 stochastic inferences for test cases L01–R08 are provided in Supplementary Figures 2 ^{(211.5KB, tif)} and 3 ^{(222KB, tif)} for models M0.25 and M0.50, respectively.

Figure 3.

Comparison of the difference maps and their associated uncertainty map

The scaled uncertainty factors (m_σ,ROI) are shown in Supplementary Table 4. The table shows that m_σ,ROI are higher for M0.25 compared to M0.50 across all ROIs. This is because MCD uncertainty is directly affected by dropout rates, which influence the standard deviation of prediction results and subsequently affect the determination of m_σ,ROI

Supplementary Table 4.

Scaling factor m_σ,ROI for each regions of interest mask for M0.50 and M0.25

Model	PTV 42.56 Gy	PTV 49.60 Gy	Heart	Ipsilateral lung	Contralateral lung
M0.25	11.90	16.20	8.59	12.75	7.74
M0.50	7.41	8.80	6.26	9.33	6.71

PTV: Planning target volume

Figure 4 presents the DVH comparison for test case R01 using the M0.25 and M0.50 models. Across both PTVs and OARs, the predicted dose distributions (solid lines) show good agreement with the clinically planned distributions (dashed lines). The uncertainty intervals (shaded regions) exhibit minimal differences between M0.50 and M0.25 concerning variability and confidence intervals, indicating the reliability and effectiveness of the uncertainty calibration method.

Figure 4.

Dose–volume histogram comparison of the predicted (solid line) and planning (dashed line) for test case R01 with uncertainty interval range (shaded) in planning target volume s and organs at risks from M0.25 (left) and M0.50 (right) prediction

Coverage probability reflects the likelihood that the true dose lies within the model’s predictive uncertainty interval. Supplementary Table 5 reports these values for both M0.25 and M0.50 across all test cases and ROIs. For the overall body, mean coverage probabilities were 0.90 ± 0.05 for M0.25 and 0.92 ± 0.03 for M0.50. When analyzing specific ROIs (PTV, heart, ipsilateral lung, and contralateral lung), both M0.25 and M0.50 yielded comparable performance. The average coverage probabilities for M0.25 were 0.88 (PTV), 0.87 (heart), 0.97 (ipsilateral lung), and 0.89 (contralateral lung). For M0.50, the corresponding averages were 0.87, 0.91, 0.96, and 0.90. These results indicate that both models maintain robust uncertainty calibration, with consistently high coverage in the lungs and slightly lower values in the PTV and heart.

Supplementary Table 5.

Coverage probability for all test cases in total body, planning target volume, heart, ipsilateral lung, and contralateral lung for M0.25 and M0.50 (Left: Left-sided breast irradiation, Right: Right-sided breast irradiation)

Test case	Body		PTV		Heart		Ipsilateral lung		Contralateral lung
	M0.25	M0.50	M0.25	M0.50	M0.25	M0.50	M0.25	M0.50	M0.25	M0.50
Left average	0.91	0.92	0.88	0.88	0.87	0.89	0.98	0.98	0.90	0.94
Left SD	0.04	0.03	0.09	0.13	0.12	0.10	0.01	0.02	0.06	0.04
Right average	0.88	0.91	0.88	0.86	0.87	0.92	0.95	0.93	0.87	0.85
Right SD	0.05	0.03	0.04	0.08	0.16	0.10	0.07	0.06	0.12	0.09
Total average	0.90	0.92	0.88	0.87	0.87	0.91	0.97	0.96	0.89	0.90
Total SD	0.05	0.03	0.07	0.11	0.13	0.10	0.05	0.05	0.09	0.08

SD: Standard deviation, PTV: Planning target volume

DISCUSSION

This study presents a preliminary single-institution investigation of a DL approach that integrates left-and right-sided breast VMAT dose prediction and MCD-based uncertainty quantification within a unified model. Whereas previous studies have largely focused on left-sided breast cancer,[4,12,13,14,15,33] the present study evaluates a single framework applicable to both lateralities. In addition, we provide comprehensive evaluation using multiple metrics including MSE, MAE, GPR, and dose indices, addressing a gap in the literature where only 20% of DL-based dose prediction studies report GPRs.[34]

Uncertainty quantification frameworks in DL generally distinguish between aleatoric uncertainty (inherent data-level noise) and epistemic uncertainty (model-level ignorance stemming from limited or underrepresented training data). In our architecture, MCD was intentionally selected over alternative methods,[18] such as deep ensembles or specialized loss-function modifications, because of its practical and theoretical characteristics. Theoretically, MCD primarily captures epistemic uncertainty,[18] making it ideally suited for flagging underrepresented patient anatomies. Practically, MCD can be seamlessly retrofitted into existing network architectures containing dropout layers at inference time without requiring model retraining. This capability allows our uncertainty quantification strategy to be readily extended to previously published dose prediction models without the retraining or architectural modifications required by some other uncertainty quantification methods. Furthermore, MCD is widely established in the dose prediction literature,[31,35] supporting its suitability for dose prediction tasks.

Our developed models (M0.25 and M0.50) demonstrated robust predictive capabilities for 3D radiation dose distributions, achieving comparable performance for both left-and right-sided breast irradiation cases. The models showed agreement between D_plan and D_predict within the PTV, as evidenced by low MAE and MSE values. However, we observed higher prediction errors in small high-risk PTV volumes, specifically in cases L01, L10, and R04 with volumes of 11, 8, and 4 cm³, respectively. This reduced accuracy in small target volumes is consistent with known limitations of convolutional neural networks in capturing fine spatial details.[36] As illustrated in Supplementary Figure 4 ^{(102.7KB, tif)} , error metrics (MAE and MSE) peaked in small boost volumes (4-11 cm³) before dropping and stabilizing across larger target volumes. In addition, sample size constraints likely exacerbated this effect, as boost cases represented only 19 of the 68 plans, limiting model exposure during training. The increased errors near PTV boundaries are consistent with the findings of Bai et al.[12] The dose indices analysis confirmed no significant differences between D_plan and D_predict, further validating the clinical acceptability of our approach.

In breast cancer irradiation, the heart is considered a high-priority OAR because of its clinical significance.[37] The ipsilateral lung, another important OAR, exhibited moderate prediction errors. These results demonstrate the ability of the proposed models to achieve accurate dose predictions in clinically important structures, particularly those prioritized in the treatment planning process. In contrast, the contralateral lung presents the highest prediction errors among all OARs, with notable differences observed in the contralateral lung indices for the right-sided breast irradiation cases. This finding likely reflects the lower priority assigned to the contralateral lung during treatment planning optimization. Consequently, the contralateral lung experiences greater dose variation between treatment plans, making accurate prediction more challenging for the neural network.

While direct comparison remains challenging due to differences in treatment protocols, patient datasets, and planning parameters across studies,[12,38] our GPR analysis demonstrates competitive performance with existing literature. Our models achieved GPR values of 0.86 ± 0.06 (3%/3 mm) and 0.70 ± 0.08 (2%/2 mm), which compare favorably with the limited published data on VMAT breast cancer dose prediction. Specifically, Bai et al.[12] reported a GPR of 0.85 ± 0.3 (3%/3 mm), while Ahn et al.[4] achieved an average 2D GPR of 0.85 (3%/3 mm) and 0.71 (2%/2 mm). Studies using 3D conformal radiation therapy have reported slightly higher GPR values (0.89 ± 0.04 for 3%/3 mm),[14] though the differences in treatment technique complexity limit this comparison.

DI comparisons further support the accuracy of our models. Our average percentage differences in mean dose were 0.54%,-0.31%, and 1.63% for PTV, heart, and ipsilateral lung, respectively, for M0.25, and-0.10%,-0.39%, and 0.53% for M0.50. These results align closely with previous findings, such as 0.7%, 0.1%, and 0.2% reported by Bakx et al.[13] Similarly, our MAE values for PTV, heart, and ipsilateral lung (2.19%, 2.04%, and 4.03% for M0.25; 2.07%, 2.48%, and 4.64% for M0.50) are consistent with Bai et al.’s[12] results (2.40%, 4.28%, and 4.16%). For the body ROI, our MAE (M0.25: 3.99% and M0.50: 4.54%) is lower than the 5.71% previously reported.[14] These results show the models achieved primary PTV accuracy metrics broadly comparable to published left-sided-only studies despite the additional complexity of bilateral coverage, which is an encouraging preliminary finding.

The MCD uncertainty maps revealed regions of high uncertainty that were primarily concentrated around areas with high absolute differences, showing a correlation of approximately 0.59 between uncertainty and prediction error. This finding aligns with results reported by Nguyen et al.,[31] where the Pearson correlation coefficient between the absolute error and the uncertainty derived from MCD was around 0.55. Elevated uncertainty was also observed in case R05, where an atypical arm position not well-represented in the training data coincided with higher prediction error, suggesting that MCD uncertainty may serve as an indicator of out-of-distribution input. Our framework quantifies uncertainty at the voxel level, similar to the approach used by Nguyen et al.[31] While voxel-level uncertainty provides physicians with localized spatial feedback to identify where prediction errors are likely to occur, image-or plan-level uncertainty yields a single global assessment of overall plan reliability. Clinically, these voxel-level uncertainties can translate into plan-level DVH confidence bands [Figure 4]. This provides clinicians with confidence bounds on essential dosimetric parameters, such as target D_95% or OARs D_mean. A previous study[35] also measured uncertainty at the image level, which may provide an automated triage mechanism during plan review. Cases with low image-level uncertainty could undergo routine review, whereas cases exceeding a predefined uncertainty threshold could be prioritized for detailed manual assessment. Future work may integrate both strategies using image-level uncertainty for automated triage and voxel-level uncertainty for detailed, localized clinical review.

As highlighted in numerous studies.[38,39,40] DL-based dose distribution prediction has demonstrated its potential as a valuable tool for clinical guidance across various treatment sites. The predicted dose distribution can serve as an approximation of possible treatment outcomes, while the derived DVHs for the PTV and OARs can assist in treatment planning optimization. In addition, the integration of uncertainty quantification with dose prediction offers significant clinical advantages for VMAT breast treatment planning. By providing uncertainty metrics alongside predicted dose distributions, clinicians may obtain confidence indicators that help determine whether the predicted dose in a particular region warrants additional review.[32,41] This may be particularly important when the uncertainty interval includes an institutional dose-constraint threshold for a critical OAR or high-risk PTV, as such a result could prompt additional review even when the point estimate satisfies the protocol. This uncertainty-aware algorithm has the potential to improve VMAT planning efficiency by reducing both planning time and inter-planner variability, while guiding prediction reliability.

The findings of this study further support the feasibility of applying a 3D U-Net model with MCD-based uncertainty quantification for breast VMAT dose prediction within a single institution. The model performance reflects institutional planning practices, including side-specific gantry angle ranges, 2–3 arc arrangements, and consistent treatment protocols represented in the training dataset. This suggests that DL-based dose prediction may be particularly useful when trained on standardized planning strategies. At the same time, the observed uncertainty emphasizes the importance of reporting prediction confidence, as treatment protocol variation may influence model behavior and intrinsic uncertainty. Although the overall cohort size is comparable to those reported in previous breast dose prediction studies, including Ahn et al.[4] and Bakx et al.,[13] the dataset remains relatively small and includes a limited number of patients with high-risk boost volumes. Consequently, the findings for PTV 49.60 Gy should be interpreted as preliminary, and the model’s performance for small boost targets requires further validation in a larger dataset. In addition, comparative evaluation of alternative uncertainty quantification methods would help identify the most robust and interpretable approach for clinical implementation. These future investigations will be important to further validate uncertainty-aware dose prediction and strengthen its potential role in artificial intelligence (AI)-assisted VMAT planning for breast cancer.

CONCLUSION

In conclusion, this preliminary study supports the feasibility of using a 3D U-Net model with MCD-based uncertainty quantification for VMAT dose prediction in both left-and right-sided breast cancer within a single institution. The model showed initial predictive capability for the primary PTV and priority OARs, although substantially higher errors were observed for small high-risk boost volumes. The generated uncertainty maps provided additional confidence information for clinical interpretation, showed a moderate correlation with prediction error, and identified an out-of-distribution case, suggesting their potential role as a quality indicator. These findings represent an important step toward more robust and interpretable AI-assisted VMAT planning in breast cancer and highlight the need for further validation using larger, multi-institutional datasets, cross-validation, and targeted strategies for high-risk PTV prediction before clinical deployment is considered.

Conflicts of interest

There are no conflicts of interest.

Supplementary Figure 1

Three-dimensional U-Net architecture

Supplementary Figure 2

Monte Carlo dropout uncertainty map in an example slide for test case L01-R08 from model M0.25 predictions

Supplementary Figure 3

Monte Carlo dropout uncertainty map in an example slide for test case L01-R08 from model M0.50 predictions

Supplementary Figure 4

Mean absolute error and mean squared error evaluated across high-risk planning target volume for models M0.25 and M0.50

Funding Statement

This work was not supported by any external funding or grants