Fully automated intensity-modulated radiotherapy plans for rectal cancer based on deep learning predictions of three-dimensional dose distributions

Introduction

Intensity-modulated radiotherapy (IMRT) is a crucial treatment modality for rectal cancer. However, the IMRT plan design is complex and time-consuming because it requires multiple iterative optimizations by the planner, thus increasing the workload and affecting planning efficiency (1).

Automated treatment planning (ATP) is a solution to the challenges posed by manual treatment planning, enabling rapid generation of high-quality and highly consistent treatment plans (2). Currently, ATP is mainly categorized as automated rule implementation and reasoning (ARIR) and modeling of prior knowledge-based planning (KBP). ARIR allows automated treatment plan generation by implementing predetermined clinical best practice templates and simulating the entire design process (3,4). In contrast, KBP relies on large databases of historical patient plans and uses a principal component analysis to build regression models that predict the dose-volume histogram (DVH) for new patients, thus providing precise guidance for developing new treatment plans (5-10).

Although using the DVH as an objective function can provide conditions for dose constraints and enrich the acquisition of clinical information, DVH-based prediction methods often have limited accuracy because they neglect the detailed three-dimensional (3D) spatial distribution of the dose (1). In contrast, algorithms that directly predict dose distributions offer more comprehensive spatial information, which is crucial for fine adjustments of isodose lines in clinical practice. Consequently, the field of ATP has shifted toward the prediction of 3D voxel dose distributions, significantly improving prediction accuracy through the application of machine learning and deep learning (DL) technologies (11-18). Although DL techniques have shown great potential for predicting radiotherapy dose distributions for patients, clinical challenges must be addressed by exploring how to efficiently translate these predictions into practical and precise treatment plans.

A combined DVH of dose distribution predictions based on DL and KBP techniques has led to certain advancements in designing automated plans. However, with this approach, the spatial details of dose distributions are sacrificed (19). Fan et al. (20) introduced the voxel dose mean squared error (MSE) as an optimization metric to bypass DVH optimization and achieve automated plans. However, as a feasibility study conducted with an open-source treatment planning system (TPS), the plans generated by this approach may not be directly deliverable on clinical treatment machines. Therefore, Sun et al. (21) integrated DL predictions with voxel dose-based hybrid optimization, enhancing the efficiency and quality of radiotherapy planning. Nevertheless, this method may be constrained by the limited accuracy of dose predictions attributable to the use of a two-dimensional (2D) U-Net DL model, and manual adjustments of DVH parameters are still required during optimization. Zhou et al. (22) utilized a 3D U-Net combined with a residual network to accurately predict IMRT doses for rectal cancer and improved the prediction accuracy by incorporating beam parameters as inputs. However, this approach increased the training complexity and required high-quality data.

Therefore, we used a 3D multitask training U-Net (3D MT-U-Net) DL model to accurately predict radiotherapy dose distributions for rectal cancer. In contrast to conventional single-task dose prediction networks, our model incorporates auxiliary learning tasks simultaneously predicting the spatial gradient and isodose contours alongside the continuous dose map. These tasks embed geometric and physical priors into the learning process, enhancing early-stage prediction stability and improving boundary fidelity. To mitigate potential interference and scaling conflicts among tasks, we employed a shared encoder with task-specific decoders, dynamically balanced losses through target normalization, and enforced consistency between dose and contour predictions. This integrated approach yields anatomically plausible and well-calibrated dose distributions without increasing inference cost. We aimed to achieve fully automated IMRT plans with improved efficiency and clinical acceptability.

Methods

Patients

Between January 2020 and December 2022, we enrolled 125 rectal cancer patients treated with IMRT. The cohort comprised 99 patients in the training set and 26 in the independent test set. An additional 15 rectal cancer patients treated by other physicians’ groups were enrolled as the external test set, with treatment administered from January to May 2025. This retrospective study of anonymous data was approved by the Ethics Committee of the Sun Yat-sen University Cancer Center (No. SL-B2024-715-01) and the requirement for individual consent for this retrospective analysis was waived. This study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments. The general clinical characteristics of the patients are presented in Table S1. Patient data included computed tomography (CT) images, contoured structures, and manual treatment plans. All planning CT scans were performed with patients in the prone position, immobilized using an Ofit frame combined with a thermoplastic mask. All images were obtained using a Brilliance CT scanner (Big Bore; Philips Medical Systems Inc., Bothell, WA, USA), with a scanning voltage of 140 kV and current of 280 mAs; both the scanning and reconstruction slice thicknesses were set to 3 mm, with a pitch of 1:1.

Target volume and organ at risk delineation

Target definitions for preoperative radiotherapy in rectal cancer were established with reference to ICRU Report 83 (23). The gross tumor volume (GTV) included the primary rectal tumor and enlarged lymph nodes as visualized on CT and magnetic resonance imaging (MRI). The clinical target volume (CTV) was subdivided as follows: CTV1 encompassed the GTV, the mesorectum at the level of tumor involvement, the presacral space, and portions of the sigmoid colon, rectal canal, and mesorectum within 2–5 cm craniocaudally. CTV2 included the entire rectum and mesorectum, internal iliac lymph nodes, partial obturator lymph nodes, presacral lymph nodes, and partial external iliac lymph nodes. The mid-inferior aspect also covered the sacral foramina, coccyx, and at least 1 cm of the posterior bladder wall. In male patients, the posterior border of the prostate and seminal vesicles was included; in female patients, the posterior vaginal wall and posterior cervix were incorporated. The superior border extended to above the bifurcation of the common iliac artery (typically at the mid-level of the fifth lumbar vertebra), including the rectosigmoid junction and the entire mesorectum. The planning target volume (PTV) was divided into PTV1 and PTV2, generated by expanding CTV1 and CTV2 uniformly by 0.6 cm in the craniocaudal, lateral, and anteroposterior directions. Organs at risk (OARs) included the bladder, rectum, colon, bilateral femoral heads, and spinal cord. The small intestine was contoured up to 3 cm superior to PTV2.

Treatment data

All contours were delineated by radiation oncologists for clinical treatment. All plans were designed using a Monaco TPS (version 5.1, ElektaAB, Stockholm, Sweden) and optimized based on direct machine parameter optimization using photon beams with 6 MV energy. The Monte Carlo algorithm was used to calculate the final dose. The prescribed doses for the plan were 50 Gy/25 fractions (PTV1) and 45 Gy/25 fractions (PTV2). The requirements were as follows: precision dose covers 95% of the PTV, and the dose to OARs was limited according to RTOG 0822 clinical trial (24). The specific plan dosimetric goals are detailed in Table S2. Upon plan completion, it was reviewed and confirmed by a radiation oncology expert to ensure that it met clinical treatment standards.

Data processing

The CT images, structures, and dose distributions were exported from the TPS in Digital Imaging and Communication in Medicine (DICOM) format. The DICOM data were processed using scientific computing packages (NumPy version 1.23.5 and Pydicom version 2.4.4). All 3D matrices of the CT images and dose distributions were cropped, normalized, and standardized to a uniform size of 128×128×64. The pixel values of the CT images were normalized to the range [0, 1], and the contoured structures were converted into binary images. Additionally, the resolution was adjusted to 1.37×1.37×3.00 mm (left-right, anterior-posterior, superior-inferior) to ensure data consistency. To enhance the generalizability of the model, various image augmentation strategies were applied, including random flipping, translation, rotation, and scaling. To improve model stability under various transformations, the probability of triggering augmentation operations was set to 0.3 (see details in Appendix 1).

Dose prediction model building and training

Model building

We designed the 3D MT-U-Net DL model (Figure 1). The 3D MT-U-Net model primarily uses a 3D convolutional neural network (CNN) to efficiently extract features and integrates multitask decoders to enhance performance. Additionally, a residual network (25) was incorporated into the traditional 3D U-Net architecture, effectively optimizing the information flow and reducing information loss during training, thereby improving learning efficiency and prediction accuracy.

Figure 1 The proposed deep learning model. (A) The structure of the 3D MT-U-Net model. (B) Schematic diagram of the residual module. 3D MT-U-Net, three-dimensional multitask training U-Net.

The encoder of the 3D MT-U-Net extracts key features from the input data (e.g., CT images). Through a series of convolutional and pooling operations, the spatial dimensions were progressively compressed, and the number of feature channels increased correspondingly. This process was designed to capture high-level abstract features from the data. The encoding module achieves efficient down-sampling through a carefully designed four-layer convolution, whereby each operation halves the input size and doubles the number of feature channels. For example, an input of 128×128×64 is transformed into an output of 4×4×2 after passing through the encoding module. The decoder uses up-sampling and deconvolution to restore spatial dimensions and reduce the number of feature channels until the final output matches the target image size (e.g., the dose distribution map). Rectified linear unit (ReLU) activation layers were applied to enhance the nonlinear processing capability of the model. Furthermore, skip connections were introduced during the decoding process, thus allowing the encoder and decoder to share key features, not only preserving rich spatial information and details but also significantly enhancing the ability of the model to represent various parts of the image and ultimately yielding precise and reliable dose prediction results.

We optimized the model using a dual-phase approach involving pretraining weight optimization and multitask training (Figure 2). This accelerated model convergence and significantly enhanced the ability of the model to predict complex treatment dose distributions. During the pretraining phase, the encoder-decoder model was initialized using the data of 87 patients, thus incorporating anatomical knowledge of multiple patients, which facilitated adaptation to new data during the fine-tuning phase and created a strong foundation for subsequent training.

Figure 2 Flowchart of dose prediction using a pretrained multitask-based structure. CT, computed tomography.

During the multitask training phase, the model was fine-tuned to accurately predict radiotherapy doses. By sharing the encoder, using personalized task-specific layers and a joint optimization strategy, we improved the prediction accuracy and efficiency. The shared encoder enhanced the model’s robustness, while multitask training mitigated overfitting by allowing regularization and promoting learning across tasks.

Model training

The model used a weight-sharing encoder coupled with three decoders that handled the dose, gradient, and isodose line maps, further boosting the generalizability of the model. The loss function comprised the following three components:

Lossall=α⋅Lossiso+β⋅Lossgra+γ⋅Lossdose[1]

where α is the weight for the isodose region loss, β is the weight for the dose gradient loss, and γ is the weight for the dose distribution loss.

For the isodose region decoder, cross-entropy loss was used. This loss measures the discrepancy between the output of the model and actual labels to reduce the accuracy of the doses according to the predicted results within a broad range (26). The cross-entropy loss was calculated as follows:

Lossiso=−1W×H∑iW×H∑CN−1Ai,C⋅log(Deciso(F)i,C)[2]

where W is the width of the input image, H is the height of the input image, i is the voxel index in the current channel, C is the index of the current channel, N is the total number of isodose region categories, A is the true label value, Dec_isois the output of the isodose region decoder, and F is the feature extracted by the encoder.

We determined the MSE loss function for the dose gradient decoder and dose distribution decoder (27). After normalization, the MSE effectively reflected the difference between true values and was used to update the model weights. These values were calculated as follows:

Lossgra=1W×H∑i=1W×H∑CN−1‖Decgra(F)i,C−yi,C‖[3]

Lossdose=1W×H∑i=1W×H∑CN−1‖Decdose(F)i,C−yi‖[4]

where W is the width of the input image, H is the height of the input image, i is the voxel index in the current channel, C is the index of the current channel, N is the total number of isodose region categories, Dec_gra is the output of the dose gradient decoder, Dec_doseis the output of the dose distribution decoder, F is the feature extracted by the encoder, and y is the true label value at the location.

Training details

We used the Adam optimizer with the base learning rate set to 3e−4 and weight decay of 1e−4. The model was trained for 600 epochs. The DL model was implemented using the PyTorch framework (https://pytorch.org/) and trained in an environment equipped with an Nvidia GeForce 2070 GPU (Nvidia, Santa Clara, CA, USA) and 8 GB of memory, utilizing eight threads.

ATP

The Mozi TPS (version 3.3, Manteia, Xiamen, China) was used during this study. By integrating optimization templates with clinical constraints, high-precision dose predictions can be directly converted to automated optimization objective functions, thus eliminating the need for manual intervention and automatically generating high-quality standardized treatment plans for rectal cancer. The optimization strategy included automated optimization based on the voxel dose and DVH objectives, direct optimization of dose distributions within the TPS, and rapid dose calculations using the Monte Carlo algorithm. The final output was a clinically usable treatment plan (Figure 3).

Figure 3 Automated plan flowchart. 3D, three-dimensional; DVH, dose-volume histogram.

Automated planning based on the voxel dose

The 3D dose distribution predicted by DL was used as the voxel-based optimization objective function. The planning optimization model comprised multiple weighted objective functions, with the optimization target being the photon intensity fluence map x. The goal was to minimize the weighted sum of the secondary loss function (F) for the sub-objectives while incorporating the practical constraint (C). The specific objective functions included the following: the predicted dose distribution optimization objective function f, which aimed to optimize the dose distribution to critical organs to match the predicted dose results; the equivalent uniform dose (EUD) objective function generalized equivalent uniform dose (gEUD), which was used to further reduce the dose to critical organs (28); the uniform prescription dose objective function g, which optimized the dose within the target volume to approach the prescribed dose value; and the dose fall-off objective function h, which optimized the dose around the target volume and to critical organs to adhere to dose fall-off rules. The formulation is expressed as follows:

MinF(x)=∑iNOARx{wif[di(X),diPred]+WigEUDgEUD[di(X)]}+∑JNPYVxwjg[dj(X),d^j]+WNTh[d(X),dNT]s.t.C[d(X)]≤0,x≥0[5]

f(d,dPred)=∫01max{d(v)−dPred(v),0}2dv[6]

gEUD(d)={(∫01dα(v)dv)1/α}2[7]

where w_iis the weight factor for the objective function; d_i(x) and d_j(x) are the dose maps optimized for the corresponding OAR and PTV regions, respectively; d_i^pred is the predicted dose output for the OAR region; d^J is the prescription dose for the PTV region; d(x) is the calculated dose; W is dose deposition matrix from the original plan for ray field information; C(d(x)) is the additional dose constraint/dose-volume constraint function; NOAR and NPTV are the numbers of critical organs and target regions, respectively; NT is the auxiliary ring for PTV margin expansion; and α is the biological parameter, which was set to 2 in rectal cancer plans.

Automated planning based on DVH parameters

Plan optimization based on DVH parameters involved extracting dose-volume statistics for the target volume and OARs from the dose prediction results. These statistics were used as continuous constraints to guide optimization. The DVH constraint was defined as follows: the volume receiving a dose greater than D1 should be less than V1. To integrate this constraint into the objective function, another dose value (D2) was sought so that V(D2) = V1 in the current DVH. The objective function was expressed as follows:

f=1Np∑iH(D2−Di)⋅H(Di−D1)(Di−D1)2+⋯H(Db−Da)={1,0,Db>DaDb≤Da[8]

where N is the number of voxels corresponding to the target volume, p is the weight factor, and H is the step function.

Performance evaluation

The accuracy of the 3D MT-U-Net model when capturing dose details was assessed using the mean absolute error (MAE) obtained by comparing the predicted and planned dose distributions and by comparing the DVH parameters of those distributions. To verify the superior performance of the 3D MT-U-Net model, we compared it with advanced models from existing research, including GAN-based Pix2Pix (29), diffusion-based DoseDiff (distance-aware diffusion model) (30), and MD-Dose (diffusion model based on the Mamba) (31). Additionally, the quality of plans generated by the two automated planning methods was comprehensively evaluated, including the organ dose DVH parameters and plan optimization times. Furthermore, the plan quality was assessed using PlanIQ^TM software (version 2.0; Sun Nuclear, Melbourne, FL, USA); the comprehensive scoring criteria are outlined in Table S3. To ensure consistency and comparability during the evaluation, all dose data were normalized based on a prescription dose of 50 Gy.

Comparison of 3D voxel dose distributions

The MAE was used to assess the absolute difference between the 3D dose distributions of the model’s predicted dose distribution and reference dose distribution. The MAE was calculated as follows:

MAE=1n∑i=1n|yi−y^i|[9]

where n is the total number of voxels, i is the index of the current voxel, y is the dose value at the current voxel position, and y^ is the dose value at the current voxel position as predicted by the model.

Comparison of the DVH dosimetric parameters

The target volume evaluation parameters included the conformity index (CI), homogeneity index (HI), dose received by 98% of the target volume (D_98%), percentage of the target volume receiving the prescription dose (V_100%), minimum dose (D_min), maximum dose (D_max), and mean dose (D_mean). CI and HI were calculated using formulas as follows:

CI=TVRITV×TVRIVRI[10]

HI=(D2%−D98%)/Dp[11]

where TV is the volume of the PTV in cm³, V_RI is the volume encompassed by the prescription dose isodose line in cm³, TV_RI is the volume of the target covered by the prescription dose isodose line in cm³, D_2% is the dose received by 2% of the PTV volume, D_98% is the dose received by 98% of the PTV volume, and D_p is the prescription dose.

Evaluation parameters of OARs

The evaluation included the D_max and dose percentage volume (V_xGy, representing the percentage or absolute volume receiving X Gy) for the small intestine and colon. For the bladder and left and right femoral heads, the D_mean and V_xGy were assessed.

Statistical analysis

The statistical analysis was performed using the software SPSS 25.0 (IBM Corp., Armonk, NY, USA). Categorical variables were analyzed using the Chi-squared test. Continuous variables were assessed using nonparametric tests. Statistical significance was set at P<0.05.

Results

Comparison of dose distributions predicted by different DL models with planned doses

Table 1 presents the MAE values of dose distributions for different DL models. Overall, our proposed method (3D MT-U-Net) consistently achieved the lowest MAEs across most target volumes (PTV1, PTV2) and OARs. Compared to the GAN-based Pix2Pix and the diffusion-based DoseDiff, and MD-Dose, the 3D MT-U-Net model demonstrated substantially improved prediction accuracy respectively, with notable reductions in MAE values for PTV1 (0.016±0.023 vs. 0.038±0.024, 0.033±0.018, and 0.017±0.039) and PTV2 (0.022±0.007 vs. 0.033±0.019, 0.038±0.008, and 0.029±0.005). For critical OARs such as the small intestine, colon, and bladder, our method also yielded the lowest MAEs (e.g., colon: 0.031±0.018), indicating superior dose conformity and better sparing of healthy tissues. Although MD-Dose showed competitive results for some structures, our approach consistently provided lower or comparable errors across all evaluated regions, underscoring its robustness and generalizability.

Comparison of dose distributions of the automated and MPs

The results of comparisons of the DVH parameters of the MP, voxel dose-based automated plan (AP_VD), and DVH-based automated plan (AP_DVH) are presented in Table 2 (independent test cohort) and Table 3 (external test cohort). Regarding the AP_VD, the dosimetric parameters for the target volumes and OARs were either superior to or comparable with those of the MP. Notably, the differences in the small bowel V_{35 Gy}, colon V_{45 Gy}, bladder D_mean, and femoral head V_{30 Gy} were significant (P<0.05). The dosimetric parameters of the AP_DVH for the target volumes and OARs were slightly better than or comparable to those of the MP. A comparison of the two automated planning methods showed that the AP_VD plans exhibited superior dosimetric parameters. Figure 4A displays the results of DVH comparisons of the MP, AP_VD, and AP_DVH for a patient with rectal cancer. Figure 4B shows the dose distributions and differences between the three plans for the same patient. The results indicated that both automated planning methods effectively protected the OARs, significantly reduced the bladder dose, included target volume coverage similar to that of the prescription dose, and had an improved CI.

Figure 4 Comparison of DVHs and dose distributions among the MP, AP_VD, and AP_DVH for radiotherapy for a patient with rectal cancer. (A) Comparison of DVHs. (B) Comparison of dose distributions. AP_DVH, DVH-based automated plan; AP_VD, voxel dose-based automated plan; DVHs, dose-volume histograms; MP, manual plan; PTV, planning target volume.

Comparison of quality scores and optimization times of the automated and MPs

The quality scores of both automated plans were either superior to or comparable with those of the MP, with no significant differences (P>0.05). The AP_VD group had the highest average quality score (81.24±6.15) (Table 4). Additionally, the optimization time of both automated plans was significantly reduced compared with that of the MP (Table S4). The average optimization times and standard deviations were 172.8±19.1, 43.6±2.4, and 45.1±1.7 seconds for the MP, AP_VD, and AP_DVH, respectively.

However, we also found that the predicted plan of one patient was significantly different from the reference MP (Figure S1), with a CI much higher than that of the MP (CI: 0.83 vs. 0.76), but the small intestinal dose exceeded 50 Gy. This also confirms that automatic planning may not necessarily meet clinical requirements and that manual modifications based on actual clinical needs are still needed.

Ablation study

Table 5 presents the mean values and standard deviations of dose distribution metrics under different module configurations. Overall, incorporating either the Isodose Line Map or the Gradient Map individually led to moderate improvements over the baseline in certain target volumes (PTV1, PTV2) and OAR metrics, including reductions in dose to the bladder and femoral heads.

Notably, the simultaneous inclusion of both the Isodose Line Map and the Gradient Dose Map yielded the lowest values across most evaluated metrics. Specifically, the combined configuration achieved the best results for PTV1 (0.016±0.023), PTV2 (0.022±0.007), colon (0.031±0.018), and bladder (0.071±0.031), indicating superior dose conformity and sparing of critical structures. Dose reductions were also observed in the femoral heads, further underscoring the advantage of the joint approach.

Discussion

This study introduced an innovative 3D MT-U-Net model with a 3D CNN to extract features and a multitask decoder to precisely predict the 3D dose distribution of IMRT plans for rectal cancer. This model enhanced the comprehensiveness of dose predictions and significantly improved the ability to capture dose distribution details. Compared with the clinical MPs, the 3D MT-U-Net model demonstrated exceptional predictive performance in terms of both the PTV1 region and globally, with a significant reduction in the MAE.

When compared to the GAN-based Pix2Pix (29) and the diffusion-based models DoseDiff (30) and MD-Dose (31), our newly introduced 3D MT-U-Net model consistently demonstrated the lowest MAEs across the majority of target volumes, namely PTV1 and PTV2, as well as OARs. These results underscore the 3D MT-U-Net model’s superior ability to capture intricate dose distributions, delivering the prescribed high-dose coverage to target regions while minimizing low-dose spill into adjacent OARs. The overall enhancement in dose prediction accuracy indicates that our methodology holds significant promise for clinical implementation, facilitating top-tier and efficient radiotherapy planning, a conclusion that aligns with the insights presented by Zhou et al. (22).

Many studies have successfully used DL methods to predict dose distributions for various tumors in the head and neck (11,16,20), chest (17), abdomen (13), and pelvis (12,32). However, dose prediction is only one step in the automatic planning process. Further refinement of the predicted results is necessary to generate executable plans and ultimately improve clinical efficiency and accuracy. Therefore, we designed clinically objective optimization templates and developed an automated planning optimization strategy focused on the voxel dose and DVH objective functions. This strategy integrated high-precision 3D dose prediction with OAR protection, thus enabling the rapid and fully automated design of rectal cancer radiotherapy plans. The entire process required no manual intervention, thereby significantly enhancing the efficiency of plan development while effectively reducing human error and providing a reliable and efficient personalized treatment plan for rectal cancer patients.

Compared with MP, the two automated optimization strategies developed in this study demonstrated only minor differences in the target DVH curves and maintained good similarity. This significant advantage is primarily attributable to the tailored planning templates specifically designed for rectal cancer treatment. By applying precise algorithms for balanced optimization, we ensured substantial improvements in the uniformity and accuracy of the target dose distribution. Furthermore, the automated optimization plans in this study significantly reduced the dose to OARs, and the DVH curves and key parameters were lower than or comparable with those of MPs. These results align with those of Sun et al. (21) and Hirotaki et al. (33). Notably, although manual optimization plans typically only account for the iteration time, our two automated optimization methods resulted in a substantial improvement in time efficiency. Considering the frequent adjustments that are required when designing MPs, the actual design time often exceeds the simple iteration time and sometimes spans several hours (21). In contrast, in this study, the automated optimization plans completed dose prediction to optimization and result output in 1–2 minutes, greatly enhancing the overall planning efficiency.

Although the analyses of dose distributions and DVH comparisons provided valuable references that allowed assessments of the quality of radiotherapy plans, they were insufficient for comprehensive and quantitative evaluations. To address this, we designed a detailed scoring system to allow a more precise and systematic quantitative evaluation of the automated plan quality. During the evaluation of the test set, the AP_VD group had an excellent comprehensive score, demonstrating that this method can automatically generate high-quality IMRT plans for rectal cancer.

For typical rectal cancer cases, radiotherapy planning is generally straightforward, and our system efficiently generates clinically acceptable plans that meet standard requirements without necessitating manual intervention. These automated plans serve as a robust foundation, allowing clinicians to rapidly obtain viable treatment plans while retaining flexibility for further manual refinement if desired. However, substantial human oversight remains essential in specific clinical scenarios: notably, (I) the presence of metastatic lesions requiring simultaneous irradiation with differing prescription doses introduces complex planning goals and trade-offs demanding expert input; and (II) palliative or non-curative treatment intents, where both prescription doses and OAR constraints deviate significantly from standard protocols, necessitating careful adaptation best achieved through manual intervention.

Our study had several limitations. First, although advanced network architectures and training strategies were employed, further model optimization is possible; future work should explore more complex network structures and efficient training algorithms to enhance prediction accuracy and generalization. Second, although automated plan generation was achieved, significant refinement is needed—as evidenced by cases where model-generated plans prioritized target coverage metrics such as PTV1 CI over critical clinical requirements, for example, small bowel sparing in situations with strict OAR constraints. This highlights a fundamental limitation: our model follows typical planning trade-offs but struggles with patient-specific constraints deviating from training patterns. Future improvements must integrate personalized optimization objectives/constraints, enhance template designs, and develop implementation methods that prioritize clinical priorities over standard optimization pathways. Third, the patient cohort size was relatively small; despite statistically significant trends, validation through larger multi-center cohorts is essential for generalization. Although the prediction model could extend to other cancer types with sufficient data (achieving comparable performance), planning automation faces greater challenges at sites with many small OARs (e.g., nasopharyngeal cancer), requiring further customization—an area flagged for future exploration.

Conclusions

The proposed 3D MT-U-Net model demonstrated exceptional performance and significant potential for predicting 3D dose distributions of IMRT plans for rectal cancer. Additionally, we successfully developed an innovative automated optimization strategy that integrates high-precision dose prediction models with optimization templates, thus enabling the rapid generation of IMRT plans for rectal cancer that meet stringent clinical standards.

Acknowledgments

None.

Footnote

Data Sharing Statement: Available at https://qims.amegroups.com/article/view/10.21037/qims-2025-168/dss

Funding: This work was jointly supported by National Natural Science Foundation of China (No. 12475360), the National Key Research and Development Program of China (No. 2022YFC2402304), Science and Technology Program of Guangzhou, China (No. 202206010180).

Conflicts of Interest: All authors have completed the ICMJE uniform disclosure form (available at https://qims.amegroups.com/article/view/10.21037/qims-2025-168/coif). Y.L., L.H., M.C., J.Z., R.C., Z.Q., X.D. and Y.P. report that this work was jointly supported by National Natural Science Foundation of China (No. 12475360), the National Key Research and Development Program of China (No. 2022YFC2402304), Science and Technology Program of Guangzhou, China (No. 202206010180). Z.L., Z.C., Y.L. and Q.Z. are employees of Manteia Technologies Co., Ltd. The authors have no conflicts of interest to declare.

Ethical Statement: The authors are accountable for all aspects of the work in ensuring that questions related to the accuracy or integrity of any part of the work are appropriately investigated and resolved. This retrospective study of anonymous data was approved by the Ethics Committee of the Sun Yat-sen University Cancer Center (No. SL-B2024-715-01) and individual consent for this retrospective analysis was waived. This study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments.

Open Access Statement: This is an Open Access article distributed in accordance with the Creative Commons Attribution-NonCommercial-NoDerivs 4.0 International License (CC BY-NC-ND 4.0), which permits the non-commercial replication and distribution of the article with the strict proviso that no changes or edits are made and the original work is properly cited (including links to both the formal publication through the relevant DOI and the license). See: https://creativecommons.org/licenses/by-nc-nd/4.0/.

References

1. Wang M, Zhang Q, Lam S, Cai J, Yang R. A Review on Application of Deep Learning Algorithms in External Beam Radiotherapy Automated Treatment Planning. Front Oncol 2020;10:580919. [Crossref] [PubMed]
2. Hussein M, Heijmen BJM, Verellen D, Nisbet A. Automation in intensity modulated radiotherapy treatment planning-a review of recent innovations. Br J Radiol 2018;91:20180270. [Crossref] [PubMed]
3. Richter A, Exner F, Bratengeier K, Polat B, Flentje M, Weick S. Impact of beam configuration on VMAT plan quality for Pinnacle(3)Auto-Planning for head and neck cases. Radiat Oncol 2019;14:12. [Crossref] [PubMed]
4. Janssen TM, Kusters M, Wang Y, Wortel G, Monshouwer R, Damen E, Petit SF. Independent knowledge-based treatment planning QA to audit Pinnacle autoplanning. Radiother Oncol 2019;133:198-204. [Crossref] [PubMed]
5. Wu B, McNutt T, Zahurak M, Simari P, Pang D, Taylor R, Sanguineti G. Fully automated simultaneous integrated boosted-intensity modulated radiation therapy treatment planning is feasible for head-and-neck cancer: a prospective clinical study. Int J Radiat Oncol Biol Phys 2012;84:e647-53. [Crossref] [PubMed]
6. Chang ATY, Hung AWM, Cheung FWK, Lee CH, Chan SH, Philips H, Cheng YT, Ng WT. Comparison of Planning Quality and Efficiency Between Conventional and Knowledge-based Algorithms in Nasopharyngeal Cancer Patients Using Intensity Modulated Radiation Therapy. Int J Radiat Oncol Biol Phys 2016;95:981-90. [Crossref] [PubMed]
7. Babier A, Boutilier JJ, Mcniven AL. Knowledge-based ATP for oropharyngeal cancer. Med Phys 2018;45:2875-83. [Crossref] [PubMed]
8. Faught AM, Olsen L, Schubert L, Rusthoven C, Castillo E, Castillo R, Zhang J, Guerrero T, Miften M, Vinogradskiy Y. Functional-guided radiotherapy using knowledge-based planning. Radiother Oncol 2018;129:494-8. [Crossref] [PubMed]
9. Fogliata A, Nicolini G, Clivio A, Vanetti E, Laksar S, Tozzi A, Scorsetti M, Cozzi L. A broad scope knowledge based model for optimization of VMAT in esophageal cancer: validation and assessment of plan quality among different treatment centers. Radiat Oncol 2015;10:220. [Crossref] [PubMed]
10. Tol JP, Delaney AR, Dahele M, Slotman BJ, Verbakel WF. Evaluation of a knowledge-based planning solution for head and neck cancer. Int J Radiat Oncol Biol Phys 2015;91:612-20. [Crossref] [PubMed]
11. Chen X, Men K, Li Y, Yi J, Dai J. A feasibility study on an automated method to generate patient-specific dose distributions for radiotherapy using deep learning. Med Phys 2019;46:56-64. [Crossref] [PubMed]
12. Nguyen D, McBeth R, Sadeghnejad Barkousaraie A, Bohara G, Shen C, Jia X, Jiang S. Incorporating human and learned domain knowledge into training deep neural networks: A differentiable dose-volume histogram and adversarial inspired framework for generating Pareto optimal dose distributions in radiation therapy. Med Phys 2020;47:837-49. [Crossref] [PubMed]
13. Guerreiro F, Seravalli E, Janssens GO, Maduro JH, Knopf AC, Langendijk JA, Raaymakers BW, Kontaxis C. Deep learning prediction of proton and photon dose distributions for paediatric abdominal tumours. Radiother Oncol 2021;156:36-42. [Crossref] [PubMed]
14. McIntosh C, Welch M, McNiven A, Jaffray DA, Purdie TG. Fully automated treatment planning for head and neck radiotherapy using a voxel-based dose prediction and dose mimicking method. Phys Med Biol 2017;62:5926-44. [Crossref] [PubMed]
15. Barragán-Montero AM, Thomas M, Defraene G, Michiels S, Haustermans K, Lee JA, Sterpin E. Deep learning dose prediction for IMRT of esophageal cancer: The effect of data quality and quantity on model performance. Phys Med 2021;83:52-63. [Crossref] [PubMed]
16. Liu Z, Fan J, Li M, Yan H, Hu Z, Huang P, Tian Y, Miao J, Dai J. A deep learning method for prediction of three-dimensional dose distribution of helical tomotherapy. Med Phys 2019;46:1972-83. [Crossref] [PubMed]
17. Barragán-Montero AM, Nguyen D, Lu W, Lin MH, Norouzi-Kandalan R, Geets X, Sterpin E, Jiang S. Three-dimensional dose prediction for lung IMRT patients with deep neural networks: robust learning from heterogeneous beam configurations. Med Phys 2019;46:3679-91. [Crossref] [PubMed]
18. Miotto R, Wang F, Wang S, Jiang X, Dudley JT. Deep learning for healthcare: review, opportunities and challenges. Brief Bioinform 2018;19:1236-46. [Crossref] [PubMed]
19. Xia X, Wang J, Li Y, Peng J, Fan J, Zhang J, Wan J, Fang Y, Zhang Z, Hu W. An Artificial Intelligence-Based Full-Process Solution for Radiotherapy: A Proof of Concept Study on Rectal Cancer. Front Oncol 2021;10:616721. [Crossref] [PubMed]
20. Fan J, Wang J, Chen Z, Hu C, Zhang Z, Hu W. Automatic treatment planning based on three-dimensional dose distribution predicted from deep learning technique. Med Phys 2019;46:370-81. [Crossref] [PubMed]
21. Sun Z, Xia X, Fan J, Zhao J, Zhang K, Wang J, Hu W. A hybrid optimization strategy for deliverable intensity-modulated radiotherapy plan generation using deep learning-based dose prediction. Med Phys 2022;49:1344-56. [Crossref] [PubMed]
22. Zhou J, Peng Z, Song Y, Chang Y, Pei X, Sheng L, Xu XG. A method of using deep learning to predict three-dimensional dose distributions for intensity modulated radiotherapy of rectal cancer. J Appl Clin Med Phys 2020;21:26-37. [Crossref] [PubMed]
23. Gregoire V, Mackie TR, de Neve W, Gospodarowicz M, Purdy JA, van Herk M, Miemierko A. ICRU Report 83: Prescribing, recording, and reporting intensity-modulated photon-beam therapy (IMRT). J ICRU 2010;10: [Crossref]
24. Zhong H, van Soest J, Wang J, Valentini V, Geng H, Huang M, Cheng C, Garofalo M, Hong TS, Lerma FA. External Validation of the European Rectal Cancer Prognosis Model Using NRG Oncology RTOG 0822 Clinical Trial Data. Int J Radiat Oncol Biol Phys 2017;99:E206.
25. Wu Z, Shen C, van den Hengel A. Wider or deeper: revisiting the resnet model for visual recognition. Pattern Recognit 2019;90:119-33.
26. Connor R, Dearle A, Claydon B, Vadicamo L. Correlations of Cross-Entropy Loss in Machine Learning. Entropy (Basel) 2024;26:491. [Crossref] [PubMed]
27. Tan S, Tang P, Peng X, Xiao J, Zu C, Wu X, Zhou J, Wang Y. Incorporating isodose lines and gradient information via multi-task learning for dose prediction in radiotherapy. In: de Bruijne M, Cattin PC, Cotin S, editors. Medical Image Computing and Computer Assisted Intervention – MICCAI 2021. Springer 2021:753-63.
28. Fogliata A, Thompson S, Stravato A, Tomatis S, Scorsetti M, Cozzi L. On the gEUD biological optimization objective for organs at risk in Photon Optimizer of Eclipse treatment planning system. J Appl Clin Med Phys 2018;19:106-14. [Crossref] [PubMed]
29. Isola P, Zhu JY, Zhou T, Efros AA. Image-to-Image Translation with Conditional Adversarial Networks. 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), Honolulu, HI, USA, 2017:5967-76, doi: 10.1109/CVPR.2017.632.
30. Zhang Y, Li C, Zhong L, Chen Z, Yang W, Wang X. DoseDiff: Distance-Aware Diffusion Model for Dose Prediction in Radiotherapy. IEEE Trans Med Imaging 2024;43:3621-33. [Crossref] [PubMed]
31. Fu L, Li X, Cai X, Wang X, Shen Y, Yao Y. MD-Dose: A diffusion model based on the Mamba for radiation dose prediction. 2024 IEEE International Conference on Bioinformatics and Biomedicine (BIBM) 2024;911-8. doi: 10.1109/BIBM62325.2024.10822581.
32. Nguyen D, Long T, Jia X, Lu W, Gu X, Iqbal Z, Jiang S. A feasibility study for predicting optimal radiation therapy dose distributions of prostate cancer patients from patient anatomy using deep learning. Sci Rep 2019;9:1076. [Crossref] [PubMed]
33. Hirotaki K, Tomizawa K, Moriya S, Oyoshi H, Raturi V, Ito M, Sakae T. Fully automated volumetric modulated arc therapy planning for locally advanced rectal cancer: feasibility and efficiency. Radiat Oncol 2023;18:147. [Crossref] [PubMed]