Deep learning-powered radiotherapy dose prediction: clinical insights from 622 patients across multiple sites tumor at a single institution

Abstract

Purpose

Accurate pre-treatment dose prediction is essential for efficient radiotherapy planning. Although deep learning models have advanced automated dose distribution, comprehensive multi-tumor analyses remain scarce. This study assesses deep learning models for dose prediction across diverse tumor types, combining objective and subjective evaluation methods.

Methods and materials

We included 622 patients with planning data across various tumor sites: nasopharyngeal carcinoma (n = 29), esophageal carcinoma (n = 82), left-sided breast carcinoma (n = 107), right-sided breast carcinoma (n = 95), cervical carcinoma treated with radical radiotherapy (n = 84), postoperative cervical carcinoma (n = 122), and rectal carcinoma (n = 103). Dose predictions were generated using U-Net, Flex-Net, and Highres-Net models, with data split into training (60%), validation (20%), and testing (20%) sets. Quantitative comparisons used normalized dose difference (NDD) and dose-volume histogram (DVH) metrics, and qualitative assessments by radiation oncologists were performed on the testing set.

Results

Predicted and clinical doses correlated well, with NDD values under 3% for tumor targets in nasopharyngeal, breast, and postoperative cervical cancer. Qualitative assessments revealed that U-Net, Flex-Net, and Highres-Net achieved the highest accuracy in cervical radical, breast/rectal/postoperative cervical, and nasopharyngeal/esophageal cancers, respectively. Among the test cases (n = 123), 53.7% were deemed clinically acceptable and 32.5% required minor adjustments. The “Best Selection” approach, combining strengths of all three models, raised clinical acceptance to 62.6%.

Conclusion

This study demonstrates that automated dose prediction can provide a robust starting point for rapid plan generation. Leveraging model-specific strengths through the “Best Selection” approach enhances prediction accuracy and shows potential to improve clinical efficiency across multiple tumor types.

Supplementary Information

The online version contains supplementary material available at 10.1186/s13014-025-02634-7.

Background

Radiotherapy (RT) is a critical treatment modality for tumors [1]. Among available techniques, Intensity-Modulated Radiation Therapy (IMRT) has become the preferred method for targeting solid tumors with precision, minimizing exposure to vital organs and surrounding tissues. Volumetric Modulated Arc Therapy (VMAT) [2], an advanced form of IMRT, employs rotating arcs for superior modulation capabilities, allowing for greater conformity to tumor shape and location. This precision enhances dose delivery while better sparing healthy tissues. However, planning VMAT treatments is time-intensive, often requiring iterative adjustments by experienced dosimetrists.

Key time-consuming steps in the planning process include contouring plan-assisted structures and adjusting parameters [3], both of which demand significant time and expertise to optimize. In this context, having predictive information about dose distribution in advance could streamline the process by allowing optimization based on limited plan-assisted contours, potentially saving time. Additionally, predictive dose distribution could offer valuable visual input and dose-volume metrics, supporting physicians’ decision-making before initiating the optimization process [4].

Recent advances in deep learning, especially convolutional neural networks (CNNs), have shown promising potential in radiation dose prediction [2–22]. These methods exceed the accuracy and efficiency of conventional knowledge-based planning (KBP) techniques [23, 24]. Studies have highlighted the predictive capabilities of deep learning models across diverse cancer types, including head and neck [7–12], esophageal [5, 6], breast [13, 14], lung [15], pancreatic [16], cervical [17–21], prostate [2–4, 15, 16], and rectal cancer [22]. Nevertheless, despite these advances, existing literature lacks a comprehensive evaluation of deep learning models’ effectiveness across multiple tumor types.

This study addresses this gap by assessing the effectiveness of deep learning models in predicting radiation doses across multiple tumor sites. Specifically, we conduct a detailed quantitative and qualitative evaluation of three distinct model architectures.

Materials and methods

Patient characteristics and imaging details

The study was approved by the Ethics Committee of Nanjing Drum Tower Hospital, and informed consent was waived due to the retrospective nature of the study, which posed no additional health risks. Planning data from patients with nasopharyngeal, esophageal, breast, cervical, and rectal carcinomas who received radiotherapy at Nanjing Drum Tower Hospital were evaluated in reverse chronological order, starting from October 8, 2021. The inclusion criteria were as follows: (a) patients with a single radiation target treated at our institution; (b) no metastases; (c) availability of planning CT scans, RT structure, and RT dose in DICOM format; and (d) all plans and doses reviewed by expert dosimetrists and subsequently approved by oncologists with over eight years of experience.

A total of 622 patients met the inclusion criteria were identified across seven datasets, including nasopharyngeal (NPC, n = 29), esophageal (EC, n = 82), left-sided breast cancer cases (n = 107) and right-sided cases (n = 95), cervical radical carcinoma (n = 84), postoperative cervical carcinoma (n = 122), and rectal carcinoma (n = 103).

The planning CT images were acquired using a Siemens SOMATOM Definition CT scanner (Siemens, Erlangen, Germany) following a standardized clinical protocol: 120 kVp, 120–300 mAs, pixel spacing of 0.8 × 0.8, image resolution of 512 × 512 pixels, and slice thickness of 5 mm.

Planning details

The planning target volume (PTV) and organs at risk (OARs) were meticulously contoured and thoroughly reviewed by experienced radiation oncologists. Volumetric Modulated Arc Therapy (VMAT) plans were designed using the Monaco planning system (Elekta, Stockholm, Sweden) using a 6MV 360-degree arc. All plans were subsequently clinically approved by the responsible senior oncologists.

Prescription doses and fractions were administered based on the specific tumor site. For patients with nasopharyngeal carcinoma (NPC), the prescribed doses were 70 Gy/64Gy/58Gy for PGTV/PTV-LN (PTV1)/PTV (PTV2), respectively, delivered over 33 fractions. Patients with esophageal cancer (EC) received 50.4 Gy over 28 fractions to the PTV. Breast carcinoma patients were treated with 50 Gy over 25 fractions to the PTV. Cervical carcinoma patients received a prescription dose of 50.4 Gy over 28 fractions to the PTV, while rectal carcinoma patients received 50.4 Gy over 28 fractions to their respective PTVs.

Deep learning models for dose prediction

As depicted in Fig. 1, three CNN-based deep learning models namely U-Net, Flex-Net, and Highres-Net, were assessed for each dataset. These models are commercially available on the Acculearning AI Model Training Platform, provided by Manteia Medical Technologies. The platform enables no-coding workflows with features such as automatic data cleansing, adjustable multi-layer CNN architectures, and model performance evaluation, allowing users to customize training parameters for dose prediction.

Fig. 1.

Workflow for automated dose prediction across multiple tumor sites. (A) Input multi-site CT images with RT contours and dose data; (B) Deep learning model pipeline; (C) Axial, sagittal, and coronal views of the predicted dose distribution; (D) Quantitative and qualitative evaluation of prediction accuracy

Training and validation

The dataset included multiple tumor types from a total of 622 patients. Data were divided into training, validation, and testing sets at a 6:2:2 ratio. The models were trained using the mean absolute error (MAE) loss function and the Adam optimizer, which was chosen for its ability to adapt the learning rate during training. An initial learning rate of 0.001 was used, with exponential decay incorporated into the optimizer. To improve model generalization without significantly affecting performance, a higher dropout rate of 0.9 was selected. This dropout rate effectively reduced overfitting on our smaller dataset, contributing to better model robustness. To enhance model robustness, data augmentation techniques were applied, including zoom ratios from 0.85 to 1.2, rotation angles of ± 15°, and gamma variations from 0.5 to 2.0. The training was conducted over 200 epochs.

Quantitative evaluation of dose distribution by dose metrics

Model performance was evaluated by comparing predicted and actual dose distributions using dose-volume-histogram (DVH)-based metrics for OARs, as well as D₉₈ and D₂ (the minimum doses covering 98% and 2% of the PTV, respectively). Additionally, the normalized dose difference (NDD) for the PTV was calculated, defined as the difference between predicted and actual doses normalized by the prescribed dose. This evaluation was conducted separately for each dataset, with emphasis on the testing set.

Qualitative evaluation of dose distribution by experienced radiation oncologists

For qualitative evaluation, predicted results from the testing sets were visually assessed by a committee of senior radiation oncologists specializing in different tumor types. All committee members have extensive experience in radiotherapy, exceeding eight years. In this study, we performed qualitative assessments of each model’s dose prediction results and used the “Best Selection” method to identify the top predictions among U-Net, Flex-Net, and Highres-Net based on qualitative scores. The “Best Selection” method does not involve creating a new model from scratch or combining multiple models. Instead, it is a process applied on each test data to identify the best possible prediction among the three models.

The oncologists graded each predicted dose distribution on a 4-point scale: grade A-acceptance (clinical acceptance), grade B-minor revision (slight optimization required), grade C-major revision (significant optimization required), and grade D-rejection (requiring re-planning).

Statistical analysis

A repeated measures analysis of variance (ANOVA) was used to compare model differences. When the assumptions of sphericity or normality for ANOVA were violated, the Friedman test, a non-parametric alternative, was employed. Paired t-tests were used to compare the predicted and actual outcomes for each model, and when the assumptions of normality or homoscedasticity were violated, the Wilcoxon signed-rank test was applied as a non-parametric alternative.

Results

Dose-volume histogram (DVH) analysis for predicted PTV doses across different tumor sites

Table 1 presents the predicted PTV results across various tumor types, highlighting performance differences among the models. For nasopharyngeal carcinoma, there was no significant difference in the D₉₈ predictions for PGTV_69.96 Gy, PTV1_60.06 Gy, and PTV2_53.46 Gy across the three models (P > 0.05). Highres-Net provided the most accurate predictions to the approved dose for the D₂ parameters of all PTVs, indicating its superior accuracy for this tumor type. In esophageal carcinoma, U-Net performed best in predicting PTV_45Gy_D₂, with minimal deviation from the approved dose (48.97 ± 1.80 Gy vs. 48.22 ± 0.47 Gy, NDD = 2.3 ± 9.87%). For left-sided breast cancer, Flex-Net had the highest accuracy (NDD = -0.29 ± 4.08%), while for right-sided breast cancer, Highres-Net’s prediction for PTV_50Gy_D₂ (52.42 ± 0.10 Gy) was closest to the approved dose (53.84 ± 0.26 Gy, NDD = 1.62 ± 4.39%). In cervical carcinoma, U-Net consistently produced predictions closely matching the approved dose for both radical and postoperative cases (PTV_50Gy_D₉₈ and _D₂). For instance, PTV_50Gy_D₉₈ predictions differed by only 0.31 Gy (49.67 Gy vs. 49.98 Gy). In rectal carcinoma, despite no significant differences among models in NDD (P = 0.595), Flex-Net showed better performance in predicting PTV_45Gy_D₉₈ and D₂, closely aligning with the approved dose (D₉₈: 45.11 Gy vs. 45.10 Gy).

Table 1.

Predicted dose-volume histogram (DVH) parameters of planning target volumes (PTVs) across multiple tumor sites for each mode

Metrics		Dose predicted by DL models							P-value
		DoseApproved		DoseU−Net	DoseFlex−Net	DoseHighres−Net
Nasopharyngeal carcinoma
PGTV69.96Gy_D98 (Gy)		67.28 ± 2.21		66.76 ± 2.86	66.86 ± 2.13	67.28 ± 2.21			0.568
PGTV69.96Gy_D2 (Gy)		75.22 ± 0.48		76.22 ± 0.70	74.83 ± 0.50	74.84 ± 0.43			0.012†
NDDPGTV (%)		N/A		-2.01 ± 3.72	-0.45 ± 3.51	0.01 ± 4.05			0.503
PTV160.06Gy_D98 (Gy)		61.77 ± 0.66		61.70 ± 1.06	61.01 ± 1.06	61.23 ± 0.69			0.062
PTV160.06Gy_D2 (Gy)		74.88 ± 0.51		75.74 ± 0.79	74.54 ± 0.51	74.59 ± 0.33			0.045†
NDDPTV1 (%)		N/A		-1.94 ± 6.03	-0.02 ± 4.35	2.57 ± 6.09			0.046†
PTV253.46Gy_D98 (Gy)		53.46 ± 0.56		53.09 ± 0.49	52.62 ± 0.86	52.87 ± 0.66			0.080
PTV253.46Gy_D2 (Gy)		74.12 ± 0.51		75.10 ± 0.75	73.84 ± 0.46	74.04 ± 0.47			0.049†
NDDPTV2 (%)		N/A		-1.20 ± 6.27	0.88 ± 5.72	-0.78 ± 5.95			0.053
Esophageal carcinoma
PTV45Gy_D98 (Gy)		45.10 ± 0.79		44.81 ± 1.02	44.99 ± 0.70	42.72 ± 0.72**			< 0.001†
PTV45Gy_D2 (Gy)		48.22 ± 0.47		48.97 ± 1.80	50.03 ± 1.81**	47.19 ± 0.25**			< 0.001†
NDD (%)		N/A		2.30 ± 9.87	3.41 ± 8.85	7.4 ± 7.71			0.002†
Breast left−sided
PTV50Gy_D98 (Gy)		49.61 ± 0.63		48.64 ± 0.38**	48.89 ± 0.22**	48.51 ± 0.28**			< 0.001†
PTV50Gy_D2 (Gy)		53.77 ± 0.19		52.16 ± 0.038**	52.45 ± 0.23**	52.46 ± 0.10**			< 0.001†
NDD (%)		N/A		1.79 ± 4.52	-0.29 ± 4.08	1.96 ± 4.33			< 0.001†
Breast right−sided
PTV50Gy_D98 (Gy)		49.75 ± 0.63		48.81 ± 0.24**	48.93 ± 0.21**	48.36 ± 0.21**			< 0.01†
PTV50Gy_D2 (Gy)		53.84 ± 0.26		52.39 ± 0.20**	52.36 ± 0.23**	52.42 ± 0.10**			0.422
NDD (%)		N/A		-1.72 ± 4.98	-2.13 ± 4.37	1.62 ± 4.39			< 0.001†
Cervical radical
PTV50Gy_D98 (Gy)		49.98 ± 0.48		49.67 ± 0.28**	49.39 ± 0.35**	48.49 ± 0.13**			< 0.001†
PTV50Gy_D2 (Gy)		53.74 ± 0.25		53.09 ± 0.40**	52.69 ± 0.45**	52.42 ± 0.07**			< 0.001†
NDD (%)		N/A		8.09 ± 22.60	9.45 ± 22.10	8.80 ± 22.28			0.223
Cervical postoperative
PTV50Gy_D98 (Gy)		49.76 ± 0.48		49.42 ± 0.36**	49.29 ± 0.33**	48.61 ± 0.18**			< 0.001†
PTV50Gy_D2 (Gy)		53.77 ± 0.18		53.05 ± 0.17**	52.72 ± 0.32**	52.51 ± 0.06**			< 0.001†
NDD (%)		N/A		-1.24 ± 6.03	0.56 ± 6.81	0.05 ± 6.41			0.136
Rectal carcinoma
PTV45Gy_D98 (Gy)		45.10 ± 0.37		44.67 ± 0.29**	45.11 ± 0.21	44.38 ± 0.19**			< 0.001†
PTV45Gy_D2 (Gy)		48.34 ± 0.18		47.41 ± 0.31**	47.71 ± 0.18**	47.25 ± 0.15**			< 0.001†
NDD (%)		N/A		5.35 ± 20.42	5.29 ± 21.56	5.97 ± 20.30			0.595

^*indicate a statistically significant difference (P < 0.05) compared to the approved dose; ^**indicate a highly significant difference (P < 0.01) compared to the approved dose; ^†indicate a significant difference among the three models; D₉₈, dose received by 98% of the volume; D₂, dose received by 2% of the volume; NDD, normalized dose difference; Breast_left−sided, breast-conserving surgery for left-sided; Breast_{right−sided}, breast-conserving surgery for right-sided; Cervical_radical, cervical carcinoma before surgery; Cervical_{postoperative}, cervical carcinoma after surgery

Figures 2, 3 and 4 illustrates the dose agreement between model predictions and the approved dose across different carcinoma types, showcasing typical examples from the test sets.

Fig. 2.

Comparison of clinical versus predicted dose distributions for representative cases of nasopharyngeal and esophageal carcinoma. The columns show: (1) clinical dose, (2) predicted dose, (3) dose difference map between clinical and predicted doses, and (4) dose-volume histogram (DVH) of the cases

Fig. 3.

Comparison of clinical versus predicted dose distributions for representative cases of breast carcinoma. The columns show: (1) clinical dose, (2) predicted dose, (3) dose difference map between clinical and predicted doses, and (4) dose-volume histogram (DVH) of the cases

Fig. 4.

Comparison of clinical versus predicted dose distributions for representative cases of cervical and ractal carcinoma. The columns show: (1) clinical dose, (2) predicted dose, (3) dose difference map between clinical and predicted doses, and (4) dose-volume histogram (DVH) of the cases

Predicted DVH parameters of OARs by different deep learning models

Table 2 compares the predicted DVH parameters for organs at risk (OARs) in nasopharyngeal, esophageal, and breast carcinoma across the three models. For nasopharyngeal carcinoma, all models provided similar predictions for critical OARs (brain stem, spinal cord, optic nerves) (P > 0.05), but Flex-Net overestimated the D_max of the left lens (6.18 ± 1.36 Gy vs. 4.56 ± 1.35 Gy, P < 0.05). For the parotid glands, all models overestimated the D_mean and V₃₀, showing highly significant differences (P < 0.05 or < 0.01). In esophageal cancer, no significant differences were found for spinalcord_D_max (P = 0.188) and lung V₅ (P = 0.087), V₂₀ (P = 0.466), and D_mean (P = 0.318) predictions. However, V₂₅ predictions for the heart differed significantly among models (P = 0.003), with Flex-Net and Highres-Net overestimating this parameter. For left-sided breast cancer, U-Net and Flex-Net significantly overestimated spinalcord_D_max (both P < 0.01). In right-sided breast cancer, Highres-Net provided the most accurate lung_V₂₀ prediction (26.17%) compared to the approved dose (26.77%).

Table 2.

Predicted DVH parameters of organs-at-risk (OARs) for nasopharyngeal, esophageal, and breast carcinomas across the three models

Metrics		Dose predicted by DL models							P-value
		Dose Approved		Dose U−Net	Dose Flex−Net	Dose Highres−Net
Nasopharyngeal carcinoma
Brain Stem
Dmax (Gy)		49.17 ± 5.68		52.68 ± 7.18	50.63 ± 6.56	50.66 ± 6.71			0.232
Spinal cord
Dmax (Gy)		38.46 ± 4.20		41.59 ± 6.34	41.65 ± 7.08	40.71 ± 6.26			0.714
Len_L
Dmax (Gy)		4.56 ± 1.35		4.54 ± 0.51	6.18 ± 1.36*	3.87 ± 1.09			0.041†
Len_R
Dmax (Gy)		4.26 ± 1.48		4.19 ± 0.48	5.09 ± 0.67	5.30 ± 0.89			0.088
OpticNerve_L
Dmax (Gy)		34.69 ± 13.57		23.92 ± 12.46	23.38 ± 13.57	21.96 ± 12.69			0.429
OpticNerve_R
Dmax (Gy)		36.48 ± 12.37		22.91 ± 7.63*	24.36 ± 8.67	24.96 ± 10.57			0.516
OpticChiasm
Dmax (Gy)		36.07 ± 17.42		25.29 ± 12.62	25.74 ± 15.37	23.80 ± 15.96			0.452
Parotid_L
Dmean (Gy)		37.59 ± 1.77		39.99 ± 3.02*	39.94 ± 2.07**	40.66 ± 2.57**			0.182
V30 (%)		65.12 ± 9.15		77.57 ± 9.58*	77.48 ± 5.01**	80.63 ± 7.45**			0.282
Parotid_R
Dmean (Gy)		36.95 ± 1.93		39.15 ± 2.72*	39.60 ± 2.66*	39.53 ± 2.57*			0.350
V30 (%)		60.60 ± 6.73		71.06 ± 7.19**	76.90 ± 7.71**	75.18 ± 6.55**			0.045†
OralCavity
Dmean (Gy)		30.20 ± 2.68		29.92 ± 1.41	30.20 ± 2.68	29.90 ± 1.20			0.824
Esophageal carcinoma
Spinal cord
Dmax (Gy)		26.71 ± 4.51		28.09 ± 1.45	28.38 ± 2.07	27.63 ± 1.29			0.188
Heart
Dmean (Gy)		3.03 ± 3.05		3.32 ± 3.22*	3.48 ± 3.16*	3.56 ± 3.24			0.443
V30 (%)		1.93 ± 2.65		1.86 ± 2.21	2.06 ± 2.50	1.98 ± 2.38			0.069
V25 (%)		3.06 ± 4.20		2.90 ± 3.53	3.20 ± 3.84	3.20 ± 3.88			0.003†
Lung_All
Dmean (Gy)		7.38 ± 1.80		7.53 ± 2.04	7.58 ± 2.01	7.61 ± 2.02			0.318
V5 (%)		41.72 ± 10.07		43.24 ± 1.96	44.33 ± 12.40	43.25 ± 12.75			0.087
V20 (%)		9.42 ± 2.89		9.92 ± 3.20	10.36 ± 4.20	10.77 ± 4.12			0.466
Breast left−sided
Spinal cord
Dmax (Gy)		20.44 ± 2.82		22.25 ± 1.13**	22.58 ± 2.42**	20.84 ± 1.19			< 0.001†
Heart
Dmean (Gy)		6.38 ± 0.90		6.42 ± 0.72	6.44 ± 0.75	6.27 ± 0.67			0.008†
V30 (%)		2.08 ± 0.97		2.10 ± 0.90	1.89 ± 0.92	1.72 ± 0.87*			< 0.001†
Lung_L
Dmean (Gy)		15.43 ± 1.13		15.71 ± 1.45	15.56 ± 1.49	15.37 ± 1.56			< 0.001†
V20 (%)		26.87 ± 1.98		28.36 ± 3.95*	27.83 ± 3.94	27.03 ± 4.25			< 0.001†
Breast right−sided
Spinal cord
Dmax (Gy)		20.82 ± 3.09		20.46 ± 1.34	21.38 ± 1.48	22.12 ± 1.14			< 0.001†
Heart
Dmean (Gy)		4.50 ± 0.51		4.44 ± 0.42	4.43 ± 0.36	4.39 ± 0.27			< 0.001†
Lung_R
Dmean (Gy)		15.54 ± 1.14		15.78 ± 1.27	15.67 ± 1.25	15.38 ± 1.18			< 0.001†
V20 (%)		26.77 ± 2.52		27.26 ± 3.12	27.29 ± 3.04	26.17 ± 3.14			< 0.001†

^*indicate a statistically significant difference (P < 0.05) compared to the approved dose; ^**indicate a highly significant difference (P < 0.01) compared to the approved dose; ^†indicate a significant difference among the three models; D_max, maximum dose received by the volume; D_mean, average dose received by the volume; V_x, percentage of the volume receiving a dose greater than or equal to x Gy; Breast_left−sided, breast-radical mastectomy for left-sided; Breast_{right−sided}, breast-radical mastectomy for right-sided

Table 3 presents DVH parameters for cervical and rectal carcinoma. In cervical cancer, all models underestimated bladder D_2cc (P < 0.001), while femoral head V₃₀ and D_mean were generally overestimated in radical cases (P < 0.01). In postoperative cervical cancer, all models underestimated intestine (D_2cc, V₃₀, V₄₀) and colon (V₃₀,V₄₀) dose parameters. Similarly, in rectal cancer, all models predicted lower values for intestine and colon D_2cc, V₃₀, and V₄₀, with Highres-Net providing the closest prediction of D_2cc to the approved dose. U-Net and Flex-Net performed better in predicting femoral head V30, with Flex-Net being the most consistent.

Table 3.

Predicted DVH parameters of OARs for cervical and rectal carcinomas across the three models

Metrics		Dose predicted by DL models							P-value
		Dose Approved		Dose U−Net	Dose Flex−Net	Dose Highres−Net
Cervical radical
Bladder
D2cc (Gy)		53.91 ± 0.34		53.14 ± 0.40**	52.5 ± 0.43**	52.27 ± 0.30**			< 0.001†
V30 (%)		69.96 ± 11.55		71.73 ± 10.02	71.55 ± 10.16	69.26 ± 10.79			< 0.001†
V40 (%)		56.23 ± 12.25		56.37 ± 10.01	56.00 ± 9.96	53.09 ± 10.14**			< 0.001†
Femurhead_L
V30 (%)		4.99 ± 2.29		7.85 ± 1.65**	9.15 ± 2.24**	5.55 ± 1.59			< 0.001†
Dmean (Gy)		15.50 ± 1.90		17.23 ± 1.48**	17.03 ± 1.66**	16.59 ± 0.88*			0.012†
Femurhead_R
V30 (%)		4.60 ± 2.41		8.67 ± 1.77**	7.64 ± 2.72**	7.39 ± 1.28**			0.057
Dmean (Gy)		14.33 ± 1.90		16.02 ± 1.57**	15.93 ± 1.08**	16.25 ± 0.88**			0.443
Intestine
D2cc (Gy)		52.61 ± 0.72		52.16 ± 0.63**	52.32 ± 0.81	52.33 ± 0.60			0.514
V30 (%)		13.93 ± 8.34		14.47 ± 9.46	14.20 ± 9.27	13.35 ± 9.27			< 0.001†
V40 (%)		6.36 ± 5.64		6.56 ± 6.31	6.47 ± 6.41	6.04 ± 6.03			< 0.001†
Colon
D2cc (Gy)		49.09 ± 7.70		49.95 ± 6.32*	49.35 ± 6.21	49.48 ± 6.44			< 0.001†
V30 (%)		20.19 ± 11.51		20.38 ± 11.51	20.28 ± 11.71	18.95 ± 11.40			< 0.001†
V40 (%)		10.93 ± 6.85		11.46 ± 7.95	11.41 ± 7.86	10.55 ± 7.54			< 0.001†
Rectum
D2cc (Gy)		52.39 ± 0.70		52.15 ± 0.58	52.02 ± 0.31	52.06 ± 0.45*			0.555
V40 (%)		57.61 ± 13.31		62.27 ± 13.17*	63.48 ± 13.45**	62.21 ± 14.08*			0.177
V50 (%)		23.04 ± 9.70		22.52 ± 6.73	25.56 ± 7.23	21.88 ± 6.85			< 0.001†
Cervical postoperative
Bladder
D2cc (Gy)		53.95 ± 0.27		52.75 ± 0.52**	52.70 ± 0.41**	52.69 ± 0.11**			0.478
V30 (%)		65.60 ± 9.64		65.64 ± 8.48	65.16 ± 8.93	64.46 ± 9.33			0.002†
V40 (%)		49.79 ± 9.54		48.75 ± 9.20	48.31 ± 9.51*	48.45 ± 9.78			0.087
Femurhead_L
V30 (%)		7.66 ± 5.17		6.58 ± 2.03	7.30 ± 2.77	5.77 ± 1.13			0.002†
Dmean (Gy)		15.99 ± 2.59		16.46 ± 1.62	16.44 ± 1.70	15.15 ± 1.14			< 0.001†
Femurhead_R
V30 (%)		7.63 ± 3.87		7.31 ± 2.34	7.24 ± 2.39	6.25 ± 1.69			0.001†
Dmean (Gy)		15.44 ± 2.63		16.08 ± 1.91**	15.99 ± 2.12*	15.80 ± 1.31			0.168
Intestine
D2cc (Gy)		53.05 ± 0.54		52.55 ± 0.42**	52.68 ± 0.66*	52.34 ± 0.41**			0.072
V30 (%)		12.50 ± 4.35		11.70 ± 4.71	11.55 ± 4.60	11.08 ± 4.66**			< 0.001†
V40 (%)		5.38 ± 2.46		5.33 ± 2.90	5.15 ± 2.82	5.13 ± 2.89			0.006†
Colon
D2cc (Gy)		44.43 ± 8.93		45.43 ± 8.08	45.01 ± 8.40	44.75 ± 7.51			0.001†
V30 (%)		16.12 ± 12.38		15.43 ± 11.05	15.20 ± 11.23	14.60 ± 10.57			0.001†
V40 (%)		8.66 ± 9.43		7.52 ± 7.15	7.30 ± 6.97	7.18 ± 6.70			0.038†
Rectum
D2cc (Gy)		51.82 ± 0.83		51.74 ± 0.85	52.29 ± 0.53*	52.40 ± 0.50**			< 0.001†
V40 (%)		47.39 ± 11.79		49.79 ± 11.69*	49.76 ± 10.67*	48.54 ± 10.42			0.072
V50 (%)		14.44 ± 6.09		16.42 ± 7.12*	16.19 ± 5.95*	13.02 ± 4.19*			< 0.001†
Rectal carcinoma
Bladder
D2cc (Gy)		48.28 ± 0.23		48.32 ± 0.34**	47.55 ± 0.18**	47.22 ± 0.46**			< 0.001†
V30 (%)		67.87 ± 9.30		66.59 ± 9.13	66.55 ± 9.23	66.96 ± 9.45			0.435
V40 (%)		49.45 ± 8.02		47.27 ± 8.42*	47.84 ± 8.90*	45.84 ± 8.80**			< 0.001†
Femurhead_L
Dmean		19.01 ± 1.88		18.94 ± 1.11	19.18 ± 1.12	18.80 ± 0.71			0.073
V30 (%)		2.84 ± 2.89		2.45 ± 1.37	3.01 ± 1.69	1.37 ± 0.74			< 0.001†
Femurhead_R
Dmean		18.51 ± 1.82		18.08 ± 1.21	18.82 ± 1.46	17.76 ± 0.98			< 0.001†
V30 (%)		2.86 ± 3.19		2.11 ± 1.44	2.91 ± 2.11	1.49 ± 1.01			< 0.001†
Intestine
D2cc (Gy)		44.81 ± 5.44		44.39 ± 5.92*	44.35 ± 6.00	44.53 ± 6.06			0.041
V30 (%)		12.72 ± 7.68		11.67 ± 7.27	11.78 ± 7.34	11.82 ± 7.57			0.442
V40 (%)		6.32 ± 5.61		5.55 ± 5.06*	5.61 ± 5.15*	5.41 ± 5.08**			0.052
Colon
D2cc (Gy)		44.06 ± 6.51		43.80 ± 6.81	43.77 ± 6.47**	43.90 ± 6.77			0.404
V30 (%)		21.59 ± 14.49		20.56 ± 13.68	21.36 ± 14.32	20.35 ± 14.30			0.039†
V40 (%)		13.43 ± 11.72		12.70 ± 11.77	12.71 ± 11.56	12.49 ± 11.60			0.195

^*indicate a statistically significant difference (P < 0.05) compared to the approved dose; ^**indicate a highly significant difference (P < 0.01) compared to the approved dose; ^†indicate a significant difference among the three models; D_mean, average dose received by the volume; V_x, percentage of the volume receiving a dose greater than or equal to x Gy; D_2cc, minimum dose received by the 2 cc hottest volume within the region; Cervical_radical, cervical carcinoma before surgery; Cervical_{postoperative}, cervical carcinoma after surgery

Qualitative evaluation

Table 4 and Fig. 5 provide qualitative scores for the three models across different tumor types. For nasopharyngeal carcinoma, Highres-Net achieved the highest percentage of grade-A (33%) and grade-B (50%) scores, while U-Net had higher proportions of lower grades, with 33% grade-D scores. Highres-Net produced no grade-D predictions, indicating its superior reliability for this tumor type. For esophageal carcinoma, Flex-Net and Highres-Net performed similarly, with Flex-Net achieving more grade-B predictions (33%) and Highres-Net performing slightly better in grade-A (20%). U-Net showed moderate performance, with 47% of predictions rated as grade-B but only 13% as grade-A. For both left- and right-sided breast cancer, Flex-Net consistently outperformed the other models in grade-A predictions (38%), while U-Net had lower scores (19% for left-sided and 16% for right-sided). Highres-Net performed better in the grade-B category (for left-sided) but showed fewer grade-A predictions. In cervical cancer (both radical and postoperative), U-Net and Flex-Net clearly outperformed Highres-Net. U-Net achieved 69% grade-A predictions in radical cases, followed by Flex-Net at 44%, while Highres-Net had only 6% grade-A predictions. In postoperative cases, U-Net and Flex-Net performed exceptionally well, with 72% and 76% grade-A scores, respectively, compared to Highres-Net’s 20%. For rectal carcinoma, Flex-Net showed the best performance, with 76% of predictions rated as grade-A, outperforming U-Net (57%) and Highres-Net (19%). Flex-Net had no grade-C or -D predictions, demonstrating its reliability for this tumor type.

Table 4.

Qualitative scores assigned to predicted dose distributions from each model

Tumor sites (testing set, n = 123 )		Qualitative Scores (A/B/C/D)
		Grade A	Grade B	Grade C	Grade D
Nasopharyngeal carcinoma (n = 6)
U-Net		1 (17%)	1 (17%)	2 (33%)	2 (33%)
Flex-Net		1 (17%)	1 (17%)	3 (49%)	1 (17%)
Highres-Net		2 (33%)	3 (50%)	1 (17%)	0 (0%)
Best Selection		2 (33%)	3 (50%)	1 (17%)	0 (0%)
Esophageal carcinoma (n = 15)
U-Net		2 (13%)	7 (47%)	5 (33%)	1 (7%)
Flex-Net		2 (13%)	5 (33%)	3 (21%)	5 (33%)
Highres-Net		3 (20%)	3 (20%)	4 (27%)	5 (33%)
Best Selection		5 (33%)	4 (27%)	5 (33%)	1 (7%)
Breastleft−sided (n = 21)
U-Net		4 (19%)	12 (57%)	5 (24%)	0 (0%)
Flex-Net		8 (38%)	12 (57%)	1 (5%)	0 (0%)
Highres-Net		1 (5%)	13 (62%)	6 (28%)	1 (5%)
Best Selection		8 (38%)	12 (57%)	1 (5%)	0 (0%)
Breastright−sided (n = 19)
U-Net		3 (16%)	11 (58%)	5 (26%)	0 (0%)
Flex-Net		7 (37%)	8 (42%)	4 (21%)	0 (0%)
Highres-Net		1 (5%)	7 (37%)	11 (58%)	0 (0%)
Best Selection		7 (37%)	8 (42%)	4 (21%)	0 (0%)
Cervicalradical (n = 16)
U-Net		11 (69%)	3 (19%)	2 (12%)	0 (0%)
Flex-Net		7 (44%)	8 (50%)	1 (6%)	0 (0%)
Highres-Net		1 (6%)	14 (88%)	1 (6%)	0 (0%)
Best Selection		13 (81%)	3 (19%)	0 (0%)	0 (0%)
Cervicalpostoperative (n = 25)
U-Net		18 (72%)	7 (28%)	0 (0%)	0 (%)
Flex-Net		19 (76%)	6 (24%)	0 (0%)	0 (%)
Highres-Net		5 (20%)	20 (80%)	0 (0%)	0 (%)
Best Selection		23 (92%)	2 (8%)	0 (0%)	0 (%)
Rectal carcinoma (n = 21)
U-Net		12 (57%)	8 (38%)	1 (5%)	0 (%)
Flex-Net		16 (76%)	5 (24%)	0 (%)	0 (%)
Highres-Net		4 (19%)	16 (76%)	1 (5%)	0 (%)
Best Selection		19 (90%)	2 (10%)	0 (%)	0 (%)

Best Selection, chosen as the best prediction from U-Net, Flex-Net, and Highres-Net based on qualitative scores (e.g., ABB→A); Breast_left−sided, breast-radical mastectomy for left-sided; Breast_{right−sided}, breast-radical mastectomy for right-sided; Cervical_preoperative, cervical carcinoma before surgery; Cervical_{postoperative}, cervical carcinoma after surgery

Fig. 5.

Stacked histogram showing the distribution of qualitative scores for each carcinoma type across different models

The “Best Selection” method, which selects the top prediction from U-Net, Flex-Net, and Highres-Net based on qualitative scores, consistently outperformed individual models across all tumor types. This method combines the best aspects of each model, offering more reliable and optimal prediction results compared to using any single model alone.

Discussion

This study systematically compared the performance of three deep learning models, namely U-Net, Flex-Net, and Highres-Net, in predicting radiotherapy dose distributions across various tumor types, including nasopharyngeal, esophageal, breast, cervical, and rectal carcinomas. For nasopharyngeal, breast, and postoperative cervical cancer, the average normalized dose difference (NDD) within the tumor targets was below 3%. Qualitative evaluation indicated that Flex-Net achieved superior accuracy in breast cancer dose predictions, while U-Net performed best for cervical carcinoma (radical). Highres-Net was more effective for nasopharyngeal carcinoma. Additionally, the “Best Selection” method (Table 4; Fig. 5) proved to be a highly reliable approach, consistently yielding the best qualitative outcomes by selecting the top-performing prediction from the three models.

In recent years, deep learning techniques have gained significant traction in the field of radiation oncology [25], presenting innovative approaches for automating dose distribution predictions. These models have already demonstrated successful applications across various tumor sites [2–22], showcasing their potential to enhance treatment precision. However, to date, no studies have conducted multi-model dose prediction research across different tumor types. This is primarily due to the time-consuming nature of data preprocessing, model training, and code development, which has presented challenges for radiation oncologists and medical physicists in advancing this area of research. With the increasing availability of commercial software platforms such as the AccuLearning platform used in this study, it is now possible to conduct various studies with higher efficiency and minimal coding effort [26]. These advancements enable tasks such as automatic segmentation of target volumes and organs at risk [26–29], as well as the automated dose prediction utilized in this study. In previous studies on nasopharyngeal carcinoma [10], esophageal cancer [5, 6], breast cancer [14, 30], cervical cancer [17–19], and rectal cancer [3, 22], model prediction results were typically evaluated only through dosimetric parameters, without incorporating subjective qualitative analysis. However, considering that the dose distributions in radiotherapy plans used for training lack a definitive gold standard, subjective evaluation of the prediction results is essential [31]. Such qualitative analysis provides a more comprehensive assessment of model performance, especially when dosimetric parameters may appear similar, but the clinical applicability of the predictions could vary. By integrating both quantitative and qualitative analyses, this approach offers a more accurate basis for model optimization and strengthens the foundation for their application in clinical practice.

In this study, we trained and evaluated three deep learning models using data from 622 patients across five distinct carcinoma types, representing a diverse range of target shapes, sizes, and anatomical locations. Our quantitative and qualitative analyses underscore the tumor-specific strengths of different dose prediction models. Quantitatively, as shown in Tables 1, 2 and 3, Highres-Net demonstrated superior performance in anatomically complex regions such as nasopharyngeal carcinoma, accurately predicting dose-volume histogram (DVH) parameters with minimal deviation from the approved dose. In contrast, Flex-Net consistently performed excelled in breast and rectal cancers, likely due to the simpler anatomical structures in these regions. This trend was corroborated by the qualitative analysis in Table 4; Fig. 5, where Flex-Net achieved higher qualitative scores for rectal carcinoma, while Highres-Net excelled in nasopharyngeal carcinoma. These findings suggest that each model leverages specific strengths aligned with the tumor type and its associated anatomical complexity. This tumor specificity may result from the variations in target volume and organ-at-risk (OAR) geometry across cancer types, with certain model architectures better suited to these characteristics. Highres-Net’s strong performance in complex regions like the head and neck may be due to its enhanced feature extraction capabilities, while Flex-Net’s reliable predictions in breast and rectal cancers likely reflect its generalized training on simpler structures. Prior studies, such as those by Wen et al. [21] and Li et al. [13], have highlighted the utility of different models for dose prediction in cervical and breast cancers, respectively. Our proposed “Best Selection” scheme, which consolidates the strengths of all three models, consistently outperformed individual models across all tumor types, supporting the potential of a hybrid approach to achieve reliable prediction outcomes. Furthermore, patient-specific dose predictions generated by this scheme can guide further refinement of treatment plans developed from initial dose predictions. When initial fluence proves suboptimal, optimization can proceed by using these patient-specific dose predictions as target objectives, with the initial fluence serving as a baseline [12]. As a result, this approach enhances the precision and effectiveness of radiotherapy, ensuring that dose distributions are finely tuned to both the tumor and surrounding organs-at-risk.

Despite the promising performance of these deep learning models across multiple tumor types, several limitations remain. First, the models were trained on data from a single institution, suggesting a need for further training on multi-institutional datasets to enhance generalizability. Additionally, external validations are essential to confirm model robustness and accuracy. The “black-box” nature of deep learning models, such as U-Net, Flex-Net, and Highres-Net, limits their interpretability, which could hinder clinical adoption. Moreover, this study only includes convolution-based models. To offer a more comprehensive comparison, we plan to explore Transformer-based models in future research. This will provide valuable insights into the performance differences between convolutional and attention-based architectures in predicting radiotherapy dose distributions, further enriching our analysis. Finally, although the clinical doses used for model training were physician-approved, they may not represent optimal values, highlighting the need for future studies to incorporate case-specific quality assurance (QA) to enhance precision. Addressing these limitations will enhance model applicability and support the development of reliable predictive tools for clinical radiotherapy.

Conclusion

In conclusion, this study evaluated the predictive performance of three deep learning models for radiotherapy dose distribution across five carcinoma types, revealing tumor-specific strengths among the models. The “Best Selection” approach effectively combined these strengths, achieving enhanced prediction accuracy across all tumor sites. These findings underscore the potential of deep learning models to advance radiotherapy planning, minimize clinician workload, and ultimately improve treatment outcomes.

Electronic supplementary material

Below is the link to the electronic supplementary material.

Abbreviations

NPC

Nasopharyngeal Carcinoma

EC

Esophageal Carcinoma

Breast_left−sided

Left-Sided Breast Carcinoma

Breast_{right−sided}

Right-Sided Breast Carcinoma

Cervical_radical

Cervical Carcinoma Treated with Radical Radiotherapy

Cervical_{postoperative}

Cervical Carcinoma Receiving Postoperative Radiotherapy

RC

Rectal Carcinoma

DVH

Dose-Volume Histograms

RT

Radiotherapy

IMRT

Intensity-Modulated Radiation Therapy

VMAT

Volumetric Modulated Arc Therapy

CNN

Convolutional Neural Networks

KBP

Knowledge-Based Planning

MAE

Mean Absolute Error

PTV

Planning Target Volume

OAR

Organ at Risk

NDD

Normalized Dose Difference

Author contributions

Original idea and study design (SSL, JZ, ZH); Data collection and analysis (SBG, ZH, YZ, LQ, JBG, JL); Manuscript writing (ZH, SSL, JZ); Important contributions to model Training (ZH, ZDL, JBG). All authors read and approved the final manuscript.

Funding

This study was supported by the National Natural Science Foundation of China (No. 82202300, 82172072), the Foundation of Shandong Provincial Key Research and Development (No. 2024CXPT084), and the Noncommunicable Chronic Diseases-National Science and Technology Major Project (No. 2024ZD0519901).

Data availability

The datasets used and/or analyzed during the current study are available from the corresponding author on reasonable request.

Declarations

Ethics approval and consent to participate

The ethics committee approved this retrospective research at Nanjing Drum Tower Hospital, and informed consent was waived. The radiotherapy data were anonymized for the scientific purpose of this work. This retrospective study was performed in line with the principles of the Declaration of Helsinki.

Consent for publication

Not applicable.

Competing interests

The authors declare no competing interests.