Optimizing Cardiac Dose Prediction in Left-Sided Breast Cancer Radiation Therapy: Clinical Strategies for Identifying Beneficiaries of the Deep Inspiration Breath-Hold Technique

Abstract

Purpose

Deep inspiration breath-hold (DIBH) is widely used in breast cancer radiation therapy to reduce cardiac radiation exposure. However, not all patients benefit from it. This study evaluated the performance of a convolutional neural network (CNN) model to predict cardiac doses under free-breathing (FB) and DIBH conditions for left-sided breast cancer radiation therapy, aiming to identify patients most likely to benefit from DIBH based on cardiac dosimetry.

Methods and Materials

A total of 265 left-sided breast cancer patients undergoing whole-breast irradiation were included, with 200 retrospectively assigned to the training set and 65 prospectively assigned to the test set. The CNN model incorporated anatomic data, including organ structures and distance-to-target volume maps, to predict 3-dimensional dose distributions. Predicted dosimetric parameters were compared with clinical data to assess accuracy, and agreement between model-based and clinical classifications of DIBH benefit was evaluated using kappa statistics.

Results

The CNN model demonstrated high accuracy in predicting cardiac dosimetric parameters, with correlation coefficients ranging from 0.84 to 0.99 for mean dose (D_mean) and D2% in the heart, left anterior descending coronary artery, and ventricles under both FB and DIBH conditions. The model also accurately predicted dose-volume histograms for these structures, with no significant differences between clinical and predicted values. Using a classification approach based on heart D_mean in FB and its reduction via DIBH, the model correctly identified 90.8% of patients as DIBH beneficiaries (kappa value, 0.876). When applying a threshold of ΔHeart D_mean ≥1 Gy, the model identified significant DIBH benefits in 63.1% of patients, with 96.9% agreement between predicted and clinical classifications.

Conclusions

The CNN-based model provides an efficient and accurate framework for predicting cardiac dose and identifying patients most likely to benefit from DIBH. Future studies should explore its applicability across broader radiation therapy scenarios and evaluate its long-term impact on cardiac outcomes.

Introduction

Radiation therapy (RT) is a critical component of adjuvant treatment for breast cancer, effectively reducing the risks of locoregional recurrence and cancer-related mortality.¹ However, irradiation of the breast or chest wall inevitably exposes the heart to radiation, thereby increasing the risk of radiation-induced cardiac toxicity. The detrimental effects of radiation on the heart are dose-dependent, with the relative risk of acute coronary events estimated to increase between 7.4% and 16.5% per gray (Gy) of mean heart dose (MHD).^2,3 The risk of radiation-induced cardiac events begins to rise within 5 years after RT and may ultimately compromise the overall therapeutic benefits of treatment.^2,3 Consequently, minimizing cardiac radiation exposure is essential for mitigating these risks.

The deep inspiration breath-hold (DIBH) technique expands lung volume and displaces the heart away from the chest wall, altering its spatial relationship to the radiation target and reducing the mean dose to the heart and left anterior descending coronary artery (LAD) by 20% to 70% compared with free-breathing (FB).4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25 Due to its accessibility and broad applicability, DIBH is the most widely used cardioprotective technique in breast cancer RT. A predictive model from a small-scale study demonstrated that the MHD in the DIBH plan was 35% lower than that in the FB plan, reducing the mean expected years of life lost due to radiation-induced cardiac damage from 0.11 years to 0.07 years.² However, the universal implementation of DIBH remains challenging, particularly in centers with limited RT resources, primarily due to the high cost and limited availability of dedicated equipment for accurate breath-hold monitoring. Furthermore, the cardioprotective benefit of DIBH varies among individuals, largely due to anatomic and respiratory differences. Certain patients, particularly those with inherently low cardiac doses in FB or those who achieve minimal dose reduction with DIBH, may experience limited or no benefit.^9,26, 27, 28, 29, 30, 31 Thus, accurate prediction of cardiac radiation dose and identification of patients who receive high cardiac doses in FB and substantial dose reductions with DIBH can enhance the efficiency of DIBH implementation and optimize treatment outcomes.

Traditionally, cardiac dose parameters are determined through individualized RT planning, which involves direct comparison of FB and DIBH plans to assess the benefit of breath-hold techniques. However, this approach is time-intensive and heavily reliant on the planner’s expertise in defining optimal objectives and constraints. Recent advancements in knowledge-based RT planning have facilitated the prediction of 3-dimensional (3D) dose distributions,³² offering a more efficient alternative to manual planning. Among deep learning techniques, convolutional neural networks (CNNs) have demonstrated considerable promise in dose prediction,^33,34 with successful applications in automated planning for breast cancer treatment.35, 36, 37, 38, 39 Nonetheless, current models inadequately predict radiation doses to critical collateral structures, such as the coronary arteries and ventricles, beyond the heart.36, 37, 38, 39

To address these gaps, we developed a 101-layer CNN model designed to predict dose distributions based on anatomic information from organs at risk (OARs), target volumes, and out-of-field distances.³³ To enhance model performance, we integrated the minimum distance from each voxel in normal structures to the planning target volume (PTV) and incorporated this metric with anatomic data, including structural maps and computed tomography (CT) images, for training the deep learning–based dose prediction network in head and neck cancers.⁴⁰ This study evaluated the predictive accuracy of this CNN model by transfer learning methods for estimating radiation doses to the heart, coronary arteries, and ventricles in whole-breast and tumor bed simultaneous integrated boost (SIB) irradiation for left-sided breast cancer in both FB and DIBH conditions. Additionally, we validated its feasibility as a tool for rapidly identifying patients most likely to benefit from DIBH based on cardiac dose prediction.

Methods and Materials

Patient selection

This study included 265 patients with left-sided breast cancer who underwent whole-breast irradiation with a simultaneous boost to the tumor bed. The training set comprised 200 patients retrospectively enrolled between January 2018 and October 2019, while the test set included 65 patients prospectively enrolled between September 2020 and April 2021. Institutional ethics approval was obtained from the medical ethics committee, and the study was registered on clinicaltrials.gov (approval number NCT05200078). Written informed consent was obtained from all participants.

RT procedure

A total of 200 patients in the training set underwent CT simulation and treatment in FB mode. The 65 patients in the test set underwent CT simulation in both FB (FB-CT) and DIBH (DIBH-CT) modes, with subsequent treatment delivered in the DIBH state. Prior to CT simulation, radiation therapists trained test set patients in DIBH using the Active Breathing Coordinator device (Elekta AB) to ensure a minimum breath-hold duration of 35 seconds and an inspiratory volume exceeding 1000 mL.

CT imaging was performed using a Brilliance CT Big Bore scanner (Philips Healthcare) with a slice thickness of 5 mm and a 512 × 512 matrix. FB-CT scans from the training set, as well as paired FB-CT and DIBH-CT scans from the test set, were transferred to the Pinnacle treatment planning system (Philips) for contouring and planning.

Initial contours for the whole-breast clinical target volume (CTV), tumor bed, and OARs were generated using a rapid auto-contouring system (processing time, 1-2 minutes) and subsequently refined by 2 experienced radiation oncologists to ensure quality. The CTV boost was created by expanding the tumor bed, followed by an additional expansion to generate the PTV boost. The PTV was generated by applying a 5 mm isotropic expansion to the corresponding CTV, and was cropped to 5 mm beneath the skin surface, for both FB and DIBH. Identical margins were applied under both breathing conditions. To maintain contouring consistency, the heart, LAD, left ventricle (LV), and right ventricle (RV) were delineated on each slice following a published CT-based atlas.⁴¹ The LAD was contoured with a 10-mm diameter paintbrush tool to represent an internal risk volume radial margin,^42,43 and no further planning risk volume expansion was applied.

The treatment plan prescribed a total dose of 43.5 Gy in 15 daily fractions, covering 95% of the PTV for the whole breast, with an SIB of 49.5 Gy in 15 fractions to the tumor bed (PTV boost). Dose constraints for OARs adhered to the institutional protocol.⁴⁴ Each intensity modulated RT (IMRT) plan was designed using a combination of 2 optimized tangential conformal beams, delivering approximately 80% of the prescribed dose, while additional IMRT beams accounted for the remaining 20%. All plans were reviewed and approved by senior physicists and experienced radiation oncologists to ensure clinical feasibility.

In total, 200 FB plans were generated for the training set patients, while each of the 65 test set patients had 1 FB plan and 1 DIBH plan, resulting in a total of 130 plans.

Prediction model

We previously developed a CNN-based intelligent system for predicting 3D dose distributions in head and neck cancers.³³ The model’s inputs included CT images with structure overlays and distance to the PTV (DPTV) maps, where each voxel in normal structures was assigned its minimum distance to the PTV surface. The outputs were corresponding dose distribution maps. A CNN workflow for pixel-wise dose prediction had also been proposed previously.⁴⁰ We used a deep learning network based on ResNet-101 with an encoder-decoder architecture, which demonstrated good performance in our previous research. The network's initial layer, Conv1, consisted of a 7 × 7 convolution with 64 filters. Conv2 to Conv5 included 3, 4, 23, and 3 deeper bottleneck architectures (DBAs), respectively. Each DBA was composed of 2 pathways: 1 processed the input using 1 × 1 and 3 × 3 convolutions, followed by batch normalization and rectified linear unit (ReLU) activation; the other served as a skip connection. The decoder consisted of 5 transposed convolutional layers, corresponding to the encoder's convolutional layers. For this study, FB-CT scans, structures, and dose maps of training set patients were preprocessed and standardized to a consistent resolution. Structures including targets and OARs were organized to facilitate region-of-interest labeling and overlapping region extraction, thereby enhancing prediction accuracy.

The CNN was trained to automatically extract multiscale image features and establish the complex nonlinear relationships between hybrid inputs and 3D dose distributions. All layers were fully retrained (end-to-end optimization) on the FB-CT training data set of the present study. A total of 65 patients were included for evaluation. FB-CT images, corresponding hybrid inputs, and reference dose maps were used to assess model performance (FB model testing). DIBH-CT images and corresponding hybrid inputs from the same cohort were directly input into the FB-CT–trained dose prediction model to generate predicted DIBH dose maps. No network layers were frozen, and no transfer learning or fine-tuning with DIBH data was performed. A comparison was then performed between actual clinical data and model predictions for both FB and DIBH plans. Additionally, the extent of heart dose reduction achieved through DIBH was assessed.

Dosimetric comparisons and evaluation of DIBH benefit

The analyzed dosimetric parameters for the PTV and PTV boost included mean dose (D_mean), dose received by 95% of the volume (D95%), and homogeneity index (HI), calculated using the equation:

$$\text{HI}=\frac{\left(\mathrm{D}2\%-\mathrm{D}98\%\right)}{\mathrm{D}50\%}$$

where D2%, D98%, and D50% represent the doses received by 2%, 98%, and 50% of the volume, respectively.

For the heart, LAD, LV, and RV, dosimetric parameters included D_mean, D2%, and relative volumes receiving 5 to 40 Gy at 5 Gy intervals (V5-V40). Additionally, the D_mean for the left lung, right lung, and right breast, as well as V5 to V30 for the left lung, were recorded.

The benefit of the DIBH technique was evaluated based on the heart D_mean in the FB plan (heart D_{mean_FB}) and the reduction in heart D_mean in the DIBH plan (heart D_{mean_DIBH}). The difference in heart D_mean (ΔHeart D_mean) was calculated as:

$$\Delta \text{Heart}{\mathrm{D}}_{\text{mean}}=\text{heart}{\mathrm{D}}_{\text{mean}\_\text{FB}}-\text{heart}{\mathrm{D}}_{\text{mean}\_\text{DIBH}}$$

The degree of heart D_mean reduction (%ΔHeart D_mean) by DIBH was determined using the equation:

$$\%\Delta \text{Heart}{\mathrm{D}}_{\text{mean}}=\left(\text{Heart}{\mathrm{D}}_{\text{mean}\_\text{FB}}-\text{Heart}{\mathrm{D}}_{\text{mean}\_\text{DIBH}}\right)/\text{Heart}{\mathrm{D}}_{\text{mean}\_\text{FB}}\times 100\%=\Delta \text{Heart}{\mathrm{D}}_{\text{mean}}/\text{Heart}{\mathrm{D}}_{\text{mean}\_\text{FB}}\times 100\%.$$

Statistical analyses

Dosimetric parameters were expressed as mean ± standard deviation (SD). Linear correlation analysis was performed to assess the relationship between the clinical and predicted D_mean and D2% values for the heart, LAD, LV, and RV dosimetry.

For comparative analysis of dose-volume histograms between clinical and predicted values in the same patients, a paired t test was used for normally distributed data, and a paired Wilcoxon signed-rank test was used for nonnormally distributed data.

Bland-Altman analysis was performed to evaluate agreement between actual clinical ΔHeart D_mean and %ΔHeart D_mean values and their corresponding model predictions. Additionally, a kappa consistency test was used to assess the agreement between actual and predicted classifications of DIBH benefit, thereby evaluating the model’s accuracy and adaptability for clinical decision-making regarding DIBH selection.

All statistical analyses were conducted using R software v4.4.1 (http://www.r-project.org/).

Results

Prediction of cardiac dosimetric parameters

The automated planning process required approximately 1 minute to generate constraint objectives for inverse optimization. Correlation analysis of cardiac dosimetric parameters in the 65 test set patients demonstrated strong linear relationships between clinical and predicted D_mean and D2% values for the heart, LAD, LV, and RV under both FB and DIBH conditions (Fig. 1, Table 1; all P < .001).

Figure 1.

Linear correlations between predicted and clinically observed mean dose (D_mean) and dose received by 2% of the volume (D2%) values for the heart, left anterior descending coronary artery (LAD), left ventricle (LV), and right ventricle (RV) in free-breathing (FB) and deep inspiration breath-hold (DIBH) plans within the test set. (A) Mean dose in FB plans, (B) mean dose in DIBH plans, (C) D2% in FB plans, and (D) D2% in DIBH plans.

Table 1.

Correlation coefficients (R) between clinical and predicted D_mean and D2% values for the heart, LAD, LV, and RV in FB and DIBH plans

	FB plan (n = 65)		DIBH plan (n = 65)
Parameter	R	P	R	P
Heart
Dmean	0.94	<.001	0.92	<.001
D2%	0.96	<.001	0.96	<.001
LAD
Dmean	0.99	<.001	0.98	<.001
D2%	0.96	<.001	0.95	<.001
LV
Dmean	0.95	<.001	0.95	<.001
D2%	0.96	<.001	0.96	<.001
RV
Dmean	0.93	<.001	0.87	<.001
D2%	0.96	<.001	0.94	<.001

Abbreviations: D2% = dose received by 2% of the volume; DIBH, deep inspiration breath-hold; D_mean = mean dose; FB, free-breathing; LAD, left anterior descending coronary artery; LV, left ventricle; RV, right ventricle.

The mean absolute error (MAE) and root mean squared error (RMSE) for predicted D_mean and D2% values for the heart, LAD, LV, and RV in both FB and DIBH plans are summarized in Table E1. For D_mean, the prediction errors were generally low across all cardiac structures in both FB and DIBH plans, with MAE and RMSE ranging from 0.41 to 1.34 Gy and 0.52 to 1.58 Gy, respectively. For D2%, the prediction errors were higher, with MAE and RMSE ranging from 1.85 to 3.22 Gy and 2.72 to 4.40 Gy, respectively. Among all structures, the LAD showed the largest variation in D2%, suggesting that high-dose regions in small structures were more sensitive to variation.

No significant differences were observed between the clinical and predicted D_mean, D2%, or dose-volume histograms at V5 to V40 for the heart, LAD, LV, and RV (Table 2, Fig. 2). Similarly, dosimetric parameters for the PTV, PTV boost, left lung, right lung, and right breast showed no significant discrepancies between clinical and predicted values (Table 2).

Table 2.

Comparison of clinical and predicted dosimetric parameters for the FB plan and DIBH plan in the test set

	FB plan (n = 65)			DIBH plan (n = 65)
Parameter	Clinical	Predicted	P	Clinical	Predicted	P
PTV
Dmean (cGy)	4672.13 ± 43.74	4668.18 ± 44	.603	4660.46 ± 42.27	4668.72 ± 49.95	.303
D95% (cGy)	4360.33 ± 26.9	4365.25 ± 21.08	.241	4359.34 ± 22.85	4362.41 ± 17.28	.426
HI	0.20 ± 0.02	0.20 ± 0.01	.681	0.19 ± 0.02	0.19 ± 0.01	.582
PTV boost
Dmean (cGy)	5125.87 ± 33.23	5122.12 ± 42.75	.575	5125.15 ± 21.43	5128.98 ± 37.8	.472
D95% (cGy)	4957.92 ± 14.71	4961.32 ± 9.95	.123	4956.54 ± 9.24	4958.19 ± 23.23	.591
HI	0.05 ± 0.01	0.05 ± 0.01	.058	0.05 ± 0.01	0.05 ± 0.01	.053
Heart
Dmean (cGy)	403.09 ± 221.85	362.94 ± 188.48	.261	224.51 ± 128.37	217.12 ± 117.32	.733
D2% (cGy)	3069.09 ± 1253.96	3082.57 ± 1274.85	.952	1774.3 ± 1326.67	1795.1 ± 1323.54	.929
V5 (%)	15.32 ± 8.46	14.73 ± 7.54	.674	8.83 ± 6.66	8.78 ± 6.19	.964
V10 (%)	8.80 ± 5.78	7.75 ± 5.17	.276	4.35 ± 4.34	3.59 ± 3.37	.269
V15 (%)	6.81 ± 5.09	5.73 ± 4.44	.201	3.06 ± 3.45	2.41 ± 2.57	.221
V20 (%)	5.65 ± 4.64	4.79 ± 3.99	.261	2.35 ± 2.90	1.87 ± 2.15	.286
V25 (%)	4.78 ± 4.25	4.19 ± 3.71	.843	1.83 ± 2.45	1.53 ± 1.88	.444
V30 (%)	3.99 ± 3.86	3.67 ± 3.45	.618	1.39 ± 2.03	1.27 ± 1.66	.712
V35 (%)	3.07 ± 3.36	3.07 ± 3.13	.992	0.97 ± 1.56	0.99 ± 1.39	.927
V40 (%)	1.69 ± 2.65	1.77 ± 2.23	.856	0.46 ± 0.96	0.40 ± 0.72	.683
LAD
Dmean (cGy)	1933.83 ± 894.59	1913.77 ± 887.79	.897	1187.7 ± 837.42	1189.38 ± 814.65	.991
D2% (cGy)	3804.52 ± 1073.06	3849.96 ± 1086.29	.811	2938.03 ± 1397.29	3019.4 ± 1379.15	.739
V5 (%)	65.07 ± 18.31	66.86 ± 18.49	.581	51.74 ± 21.44	53.47 ± 22.37	.654
V10 (%)	53.54 ± 20.85	51.09 ± 22.56	.520	35.53 ± 24.03	33 ± 23.7	.545
V15 (%)	46.99 ± 23.12	44.16 ± 23.98	.495	28.6 ± 24.14	25.46 ± 23.02	.449
V20 (%)	42.25 ± 24.53	40.18 ± 24.83	.635	23.72 ± 23.78	21.29 ± 22.24	.548
V25 (%)	38.29 ± 25.02	37.26 ± 25.08	.814	19.9 ± 22.87	18.5 ± 21.53	.720
V30 (%)	34.12 ± 25.07	34.42 ± 25.01	.946	16.6 ± 21.71	16.06 ± 20.45	.885
V35 (%)	28.77 ± 24.3	30.8 ± 24.51	.637	13.07 ± 19.77	13.48 ± 18.99	.904
V40 (%)	19.78 ± 21.4	21.95 ± 22.33	.573	8.18 ± 15.83	7.37 ± 14.14	.759
Left ventricle
Dmean (cGy)	650.93 ± 364.74	622.6 ± 337.59	.644	346.26 ± 238.02	341.58 ± 224.95	.908
D2% (cGy)	3459.23 ± 1162.60	3490.08 ± 1201.02	.882	2095.29 ± 1408.71	2132.31 ± 1431.42	.882
V5 (%)	27.28 ± 14.49	26.31 ± 12.75	.684	15.45 ± 12.84	14.93 ± 11.22	.805
V10 (%)	16.32 ± 9.83	14.81 ± 9.33	.372	7.95 ± 8.36	6.66 ± 6.64	.332
V15 (%)	12.78 ± 8.76	10.96 ± 8.07	.219	5.63 ± 6.37	4.48 ± 5.1	.258
V20 (%)	10.73 ± 8.16	9.28 ± 7.39	.289	4.33 ± 5.33	3.51 ± 4.33	.336
V25 (%)	9.17 ± 7.60	8.17 ± 6.94	.433	3.40 ± 4.52	2.91 ± 3.83	.510
V30 (%)	7.71 ± 6.94	7.18 ± 6.51	.656	2.60 ± 3.79	2.43 ± 3.41	.789
V35 (%)	5.96 ± 6.16	6.04 ± 5.97	.942	1.81 ± 2.98	1.90 ± 2.87	.866
V40 (%)	3.36 ± 4.57	3.49 ± 4.35	.861	0.85 ± 1.88	0.76 ± 1.54	.776
Right ventricle
Dmean (cGy)	394.57 ± 273.11	372.37 ± 264.12	.638	231.82 ± 142.1	221.8 ± 139.15	.685
D2% (cGy)	2212.43 ± 1346.77	2162.06 ± 1401.82	.835	1260.94 ± 1103.4	1257.83 ± 1088.45	.987
V5 (%)	18.84 ± 14.66	17.34 ± 12.91	.538	9.41 ± 9.44	9.08 ± 8.75	.837
V10 (%)	9.09 ± 9.45	7.24 ± 7.83	.226	3.65 ± 5.45	2.63 ± 3.91	.222
V15 (%)	6.44 ± 7.95	5.07 ± 6.64	.287	2.36 ± 4.15	1.61 ± 2.76	.228
V20 (%)	4.97 ± 6.89	3.92 ± 5.68	.346	1.66 ± 3.26	1.10 ± 2.05	.240
V25 (%)	3.91 ± 6.05	3.24 ± 5.12	.493	1.21 ± 2.54	0.83 ± 1.66	.317
V30 (%)	3.03 ± 5.25	2.69 ± 4.64	.698	0.84 ± 1.86	0.63 ± 1.35	.471
V35 (%)	2.11 ± 4.35	2.12 ± 4.08	.990	0.46 ± 1.08	0.43 ± 1.01	.863
V40 (%)	1.00 ± 3.30	1.07 ± 2.67	.898	0.13 ± 0.43	0.09 ± 0.29	.608
Left lung
Dmean (cGy)	976.8 ± 177.93	982.55 ± 136.4	.837	899.08 ± 173.99	912.31 ± 138.95	.633
V5 (%)	34.96 ± 5.90	34.63 ± 3.81	.706	33.32 ± 5.68	32.24 ± 4.17	.219
V10 (%)	26.32 ± 4.95	25.91 ± 3.51	.593	24.69 ± 4.76	24.00 ± 3.49	.344
V20 (%)	20.09 ± 4.43	19.32 ± 3.24	.261	18.09 ± 4.19	17.81 ± 3.06	.655
V30 (%)	15.95 ± 4.03	16.39 ± 3.18	.499	13.84 ± 3.79	14.77 ± 3.11	.126
Right lung
Dmean (cGy)	32.32 ± 15.81	29.36 ± 16.41	.289	30.92 ± 13.33	28.98 ± 12.89	.391
Right breast
Dmean (cGy)	60.08 ± 52.72	52.43 ± 56.31	.422	66.96 ± 49.87	59.84 ± 49.37	.408

Abbreviations: D2% = dose received by 2% of the volume; D95% = dose received by 95% of the volume; DIBH, deep inspiration breath-hold; D_mean = mean dose; FB, free-breathing; HI, homogeneity index; LAD, left anterior descending coronary artery; PTV, planning target volume. V5, V10, V15, V20, V25, V30, V35, and V40 denote the percentages of the volume receiving radiation of at least 5, 10, 15, 20, 25, 30, 35, and 40 Gy, respectively.

Figure 2.

Dose-volume histograms (DVHs) of cardiac structures under free-breathing (FB) and deep inspiration breath-hold (DIBH) conditions in the test set. (A) Heart in FB plans, (B) heart in DIBH plans, (C) left anterior descending coronary artery (LAD) in FB plans, (D) LAD in DIBH plans, (E) left ventricle (LV) in FB plans, (F) LV in DIBH plans, (G) right ventricle (RV) in FB plans, and (H) RV in DIBH plans.

Prediction of cardiac D_mean reduction degree

Bland-Altman analysis demonstrated good agreement between clinical and predicted values for both ΔHeart D_mean and %ΔHeart D_mean (Fig. 3). For ΔHeart D_mean, the bias was 0.064 Gy (95% CI, –0.037 to 0.166), with limits of agreement ranging from –0.740 Gy (95% CI, –0.915 to –0.566) to 0.869 Gy (95% CI, 0.695-1.044). For %ΔHeart D_mean, the bias was –0.006 (95% CI, –0.0333 to 0.0216), with limits of agreement from –0.223 (95% CI, –0.271 to –0.176) to 0.212 (95% CI, 0.164-0.259). Expanding Bland-Altman analyses to LAD, LV, and RV is shown in Fig. E1. These findings indicated moderate variability but minimal systematic bias, supporting the model’s reliability in quantifying DIBH benefits.

Figure 3.

Bland-Altman agreement analysis comparing clinical and predicted values. (A) Δ Heart D_mean and (B) %ΔHeart D_mean.

Abbreviations: D_mean = mean dose; ΔHeart D_mean = difference in heart D_mean.

Prediction of DIBH beneficiary selection

Given the absence of a standardized consensus for cardiac dosimetry in determining DIBH benefit, we evaluated the CNN model’s performance using multiple cardiac dosimetry indicators identified in previous studies.^9,26, 27, 28, 29, 30, 31

Patients were classified into 4 groups based on heart D_{mean_FB} and %ΔHeart D_mean values. Group 1 included patients with heart D_{mean_FB} ≥6 Gy and %ΔHeart D_mean ≥20%. Group 2 included those with heart D_{mean_FB} between ≥4 and <6 Gy and %ΔHeart D_mean ≥35%. Group 3 included patients with heart D_{mean_FB} <4 Gy and %ΔHeart D_mean ≥ 50%. Group 4 represented patients with limited benefit from DIBH if none of these conditions were met. The CNN model correctly classified 90.8% of patients (59/65), yielding a kappa value of 0.876, indicating strong agreement. Misclassifications occurred in 6 cases: 3 were misclassified as either benefiting or not benefiting, while another 3 were assigned to different benefit categories (Fig. 4).

Figure 4.

Classification of deep inspiration breath-hold (DIBH) beneficiaries. Patients were classified based on heart D_mean in the FB plan (heart D_{mean_FB}) and %ΔHeart D_mean values to identify those most likely to benefit from the DIBH technique.

Abbreviations: D_mean = mean dose; FB = free-breathing; ΔHeart D_mean = difference in heart D_mean.

Because increasing heart D_mean correlates with an increased risk of radiation-induced cardiac toxicity,^2,3 we used ΔHeart D_mean ≥1 Gy as a threshold for substantial benefit. Clinically, 66.2% (43/65) of the patients met this criterion, while the model predicted 63.1% (41/65). Only 2 patients (3.1%) were misclassified, yielding 96.9% agreement and a kappa of 0.933, with sensitivity 95.3% (41/43; 95% CI, 84.5-98.7), specificity 100% (22/22; 95% CI, 85.1-100.0), positive predictive value (PPV) 100% (41/41; 95% CI, 91.4-100.0), and negative predictive value (NPV) 91.7% (22/24; 95% CI, 74.2-97.7). Given the uncertainty around the 1 Gy cutoff, we assessed near-threshold performance: for patients with clinical ΔHeart D_mean 0.5 to 1.5 Gy (n = 16), the misclassification rate was 6.3% (1/16), with 0% false positives (0/6) and 10.0% false negatives (1/10).

Based on both classification approaches (the grouping criteria and ΔHeart D_mean ≥ 1 Gy), 33.8% to 43.1% of patients were classified as deriving limited benefit from DIBH. Tanguturi et al²⁸ reported no significant difference in median treatment time between DIBH and FB (4.2 vs 3.8 minutes; P = .87), whereas Zhang et al²³ reported a median per-fraction treatment time of 4 minutes. In contrast, Macrie et al⁴⁵ reported that DIBH implementation may prolong treatment time by 3 to 5 minutes per fraction. Based on a 3- to 5-minute per-fraction increase, this would correspond to an additional 45 to 75 minutes per patient over 15 fractions of RT. By integrating heart dose prediction into patient selection, unnecessary DIBH use can be avoided in at least one-third of the patients, potentially reducing radiation therapist and linear accelerator workload by 24.8 to 41.3 hours per 100 patients.

Threshold sensitivity analysis

Using ΔHeart D_mean ≥1.5 Gy, 50.8% (33/65) of patients met the criterion clinically and the model predicted 49.2% (32/65); 3 patients (4.6%) were misclassified, yielding 95.4% agreement and a kappa of 0.91, with sensitivity 93.9% (31/33; 95% CI, 79.8-99.3), specificity 96.9% (31/32; 95% CI, 83.8-99.9), PPV 96.9% (31/32; 95% CI, 83.8-99.9), and NPV 93.9% (31/33; 95% CI, 79.8-99.3). Using ΔHeart D_mean ≥2.0 Gy, 38.5% (25/65) met the criterion clinically and the model predicted 38.5% (25/65); 6 patients (9.2%) were misclassified, yielding 90.8% agreement and a kappa of 0.80, with sensitivity 88.0% (22/25; 95% CI, 68.8-97.5), specificity 92.5% (37/40; 95% CI, 79.6-98.4), PPV 88.0% (22/25; 95% CI, 68.8-97.5), and NPV 92.5% (37/40; 95% CI, 79.6-98.4).

Discussion

This study introduces a novel approach for predicting dosimetric parameters under both FB and DIBH conditions using a CNN. The model successfully predicted doses not only to the heart, but also to the LAD, LV, and RV for IMRT plans in left-sided whole-breast irradiation with an SIB to the tumor bed. A strong agreement was observed between clinically measured and model-predicted reductions in heart D_mean from DIBH, with minimal variability and bias. Additionally, the model effectively predicted the potential benefits of DIBH, achieving high consistency between clinical and predicted classifications, demonstrating robust accuracy in identifying patients most likely to benefit from DIBH.

To our knowledge, this study represents one of the largest prospective test sets in a breast cancer dose prediction study to date. Unlike previous studies on dose prediction for breast cancer RT, which typically focus on either FB or DIBH conditions separately,35, 36, 37, 38, 39 this study used a CNN to prospectively predict dosimetric parameters for the same patient under both conditions, enabling a more comprehensive patient-specific characterization of dosimetric impact and improving generalizability across clinical scenarios. In parallel, several neural network or machine learning–based decision tools have been developed to rapidly estimate heart or lung dose, most commonly from FB planning CT, to triage patients for DIBH. These approaches include models based on MHD thresholds,^38,46 lung dose–driven classification using anatomic distances,⁴⁷ rapid MHD estimation via auto-contouring combined with RapidPlan,³⁹ and logistic regression models incorporating FB-CT–derived anatomic and clinical factors.³⁰ Collectively, these FB-CT–driven selection strategies can accelerate preliminary decision-making and reduce unnecessary DIBH-CT scans. However, some patients may still achieve limited cardiac dose reduction through DIBH. To address this limitation, prior work has explored the generation of synthetic DIBH-CT from FB-CT, enabling more direct estimation of individualized dosimetric gain without routine acquisition of dual-condition imaging.⁴⁸ Building on this concept, we envision our model primarily as a post–FB simulation triage tool integrated into routine clinical workflow. In this paradigm, patients would initially undergo standard FB-CT simulation, after which a synthetic DIBH-CT could be generated to estimate the expected dosimetric benefit of DIBH. Such an approach has the potential to improve decision-making efficiency while decreasing imaging radiation exposure, resource utilization, cost, and overall patient burden.

The CNN model effectively predicted not only heart doses, but also doses to other critical cardiac structures, including the LAD, LV, and RV, with most correlation coefficients exceeding 0.95 for both D_mean and D2% in FB and DIBH plans. Currently, no established guidelines exist for dose constraints specifically addressing the heart and its substructures in breast cancer RT. For decades, heart D_mean has been considered the most practical and relevant parameter for predicting radiation-induced heart disease. However, recent dosimetric studies suggest that heart D_mean may not fully capture the exposure to specific cardiac substructures. Evidence further indicates that the dose-volume to the LV may serve as a better parameter for predicting subclinical or clinical cardiac events,^3,49, 50, 51 whereas the dose to the LAD is prioritized over the overall cardiac dose due to its strong association with coronary stenosis in hotspot areas.^51,52 Nevertheless, only 1 study on breast cancer RT dose prediction has addressed the dose to the LAD,³⁵ and none has specifically predicted ventricular doses. In Zeverino et al’s³⁵ breast cancer RT dose prediction study, the dose to the LAD exhibited significant variation, which may be attributable to increased uncertainty in dose prediction within a few voxels located in dose-gradient regions due to the small size of the LAD and its proximity to the target. As a result, more robust metrics are necessary for accurate dose prediction in these regions. To address this challenge, we developed DPTV maps and integrated them with direct anatomic information, including structure maps and CT scans, to train a combined deep learning model.⁴⁰ This approach provided a precise distance relationship between each voxel in the normal structures and target volumes, thereby improving the accuracy of dose predictions for cardiac substructures. The successful development of this multistructure cardiac dose prediction model provides a strong foundation for its future application in more complex RT protocols, expanding its potential use in determining dose constraints for multiple cardiac structures in advanced treatment regimens.

The model demonstrated high precision in rapidly identifying patients most likely to benefit from DIBH by exhibiting strong consistency between clinical and predicted classifications. A thorough assessment of the benefits and challenges associated with the implementation of DIBH requires a systematic analysis of its effects on cardiac dose. The preselection criteria for applying the DIBH technique should include: (1) a high heart dose in the FB plan, and (2) a substantial reduction in heart dose in the DIBH plan compared with the FB plan. However, few studies have comprehensively incorporated both criteria when evaluating the potential benefits of DIBH. Some studies have focused on predicting heart dose in FB plans, suggesting that DIBH may not be necessary for patients with a heart D_{mean_FB} below 3 or 4 Gy.^26,39 Other studies have sought to determine the extent of heart dose reduction achievable with DIBH by analyzing specific individual characteristics, such as body mass index,^28,53,54 breath-hold threshold,⁵⁵ lung volume,^28,30,54,56 maximum heart distance,^29,57, 58, 59 and cardiac contact distance.^27,30,54 Although these indicators are strongly associated with cardiac dose reduction during DIBH, determining the precise cardiac dose required to guide clinical decision-making remains challenging due to their inability to fully capture the heart-to-left-breast spatial relationship and to quantify the combined effect of each factor on cardiac dose.

We accurately predicted the %ΔHeart D_mean and ΔHeart D_mean using a CNN model, which effectively identified patients who would benefit most from the DIBH technique in terms of cardiac dosimetry. In the absence of established guidelines for cardiac dose evaluation in determining the clinical value of DIBH, we adopted multiple cardiac dosimetric parameters validated by prior studies.4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25 Because the incidence of late cardiac events increases with each Gy of heart D_mean and reliable cardiac dose thresholds remain undefined, no minimum limit was imposed on heart D_{mean_FB}. Instead, in this study, heart D_{mean_FB} was stratified into 3 dosimetric levels: low (<4 Gy), medium (≥4 and <6 Gy), and high (≥6 Gy). Motivated by the observation that higher heart D_{mean_FB} corresponds to a greater need for dose reduction, the %ΔHeart D_mean threshold was progressively decreased as heart D_{mean_FB} increased (≥50%, ≥35%, and ≥20%). Our results indicated that approximately 43% of the patients were classified into group 4, with either a %ΔHeart D_mean value of <20% (below the previously reported mean reduction threshold of 25%-67%),4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25 or failure to meet the requirements for the corresponding degree of DIBH reduction for heart D_{mean_FB}, resulting in limited or no benefit from DIBH.

Furthermore, given the linear relationship between heart D_mean and cardiac event risk, some studies have proposed using ΔHeart D_mean ≥1 Gy as a criterion for determining the benefits of DIBH. Cao et al²⁷ and Koide et al⁶⁰ predicted ΔHeart D_mean by combining the cardiac contact distance with the lateral heart-to-chest distance from FB-CT, reporting that 21% and 40% of the patients, respectively, had a ΔHeart D_mean of <1 Gy. Consistent with these findings, the present study found that 34% of the patients did not meet this criterion. By identifying patients most likely to benefit from DIBH, the model facilitates personalized treatment planning, providing a scientific basis for tailoring therapy to individual patients.

This study had several limitations that should be acknowledged. First, the dose prediction in this study focused exclusively on whole-breast irradiation plans for left-sided breast cancer and did not include other RT plans, such as postmastectomy chest wall irradiation or regional lymphatic drainage irradiation. Further investigations are needed to evaluate the applicability of the proposed methods across a broader spectrum of treatment scenarios. Second, although accurate 3D dose distributions were achieved for both FB and DIBH RT plans, patients underwent 2 separate CT simulations. Advances in DIBH image prediction can help reduce the additional radiation exposure associated with multiple CT scans.⁴⁸ Third, although the model showed no systematic bias at the population level, prediction variability for individual patients remained clinically noticeable (approximately ±1 Gy, as reflected in the reported results), which may affect decision-making near clinical thresholds and motivates further model refinement. In addition, because the prospective test cohort required a breath-hold duration of ≥35 seconds and an inspiratory volume >1000 mL, the data set was enriched for patients who were DIBH-feasible. This selection may overestimate model performance and the predicted benefit of DIBH compared with an unselected, real-world screening population evaluated prior to confirmation of DIBH suitability, particularly if the tool were to be applied as a pre-DIBH triage strategy. Finally, this study primarily concentrated on predicting heart dose and identifying patients who would benefit from the DIBH technique, without addressing long-term cardiac event follow-up after treatment. Although the DIBH technique may offer significant advantages for patients with high heart doses under FB and substantial dose reductions under DIBH, its actual impact on long-term cardiac toxicity requires validation through clinical studies.

Conclusions

This study advances personalized RT planning for left-sided breast cancer by leveraging a novel CNN-based model to predict cardiac dose parameters under both FB and DIBH conditions. By accurately predicting doses to critical cardiac substructures and heart dose reduction, the model provides a rapid and reliable framework for identifying patients most likely to benefit from DIBH, optimizing treatment workflow, and enhancing the effective use of cardioprotective measures.

Disclosures

None.