A feasibility study on deep learning-based individualized 3D dose distribution prediction

Deep learning-based prediction of the dose-volume histograms for volumetric modulated arc therapy of left-sided breast cancer

BACKGROUND

The advancements in artificial intelligence and computational power have made deep learning an attractive tool for radiotherapy treatment planning. Deep learning has the potential to significantly simplify the trial-and-error process involved in inverse planning required by modern treatment techniques such as volumetric modulated arc therapy (VMAT). In this study, we explore the ability of deep learning to predict organ-at-risk (OAR) dose-volume histograms (DVHs) of left-sided breast cancer patients undergoing VMAT treatment based solely on their anatomical characteristics. The predicted DVHs could be used to derive patient-specific dose constraints and dose objectives, streamlining the treatment planning process, standardizing the quality of the plans, and personalizing the treatment planning.

PURPOSE

This study aimed to develop a deep learning-based framework for the prediction of organ-specific dose-volume histograms (DVH) based on structures delineated for left-sided breast cancer treatment.

METHODS

We used a dataset of 249 left-sided breast cancer patients treated with tangential VMAT fields. We extracted delineated structures and dose distributions for each patient and derived slice-by-slice DVHs for planning target volume (PTV) and organs-at-risk. The patients were divided into training (70%, n = 174), validation (10%, n = 24), and test (20%, n = 51) sets. Collected data were used to train a deep learning model for the prediction of the DVHs based on the delineated structures. The developed deep learning model comprised a modified DenseNet architecture followed by a recurrent neural network.

RESULTS

In the independent test set (n = 51), the point-wise differences in the slice-by-slice DVHs between the clinical and predicted DVHs were small; the mean squared errors were 3.53, 1.58, 2.28, 3.37, and 1.44 [×10-4] for PTV, heart, ipsilateral lung, contralateral lung, and contralateral breast, respectively. With the derived cumulative DVHs, the mean absolute difference ± standard deviation of mean doses between the clinical and the predicted DVH were 0.08 ± 0.04 Gy, 0.24 ± 0.22 Gy, 0.73 ± 0.46 Gy, 0.07 ± 0.06 Gy, and 0.14 ± 0.14 Gy for PTV, heart, ipsilateral lung, contralateral lung, and contralateral breast, respectively.

CONCLUSIONS

The deep learning-based approach enabled automatic and reliable prediction of the DVH based on delineated structures. The predicted DVHs could potentially serve as patient-specific clinical goals used to aid treatment planning and avoid suboptimal plans or to derive optimization objectives and constraints for automated treatment planning.

Exploring the impact of network depth on 3D U-Net-based dose prediction for cervical cancer radiotherapy

Purpose

The 3D U-Net deep neural network structure is widely employed for dose prediction in radiotherapy. However, the attention to the network depth and its impact on the accuracy and robustness of dose prediction remains inadequate.

Methods

92 cervical cancer patients who underwent Volumetric Modulated Arc Therapy (VMAT) are geometrically augmented to investigate the effects of network depth on dose prediction by training and testing three different 3D U-Net structures with depths of 3, 4, and 5.

Results

For planning target volume (PTV), the differences between predicted and true values of D98, D99, and Homogeneity were statistically 1.00 ± 0.23, 0.32 ± 0.72, and -0.02 ± 0.02 for the model with a depth of 5, respectively. Compared to the other two models, these parameters were also better. For most of the organs at risk, the mean and maximum differences between the predicted values and the true values for the model with a depth of 5 were better than for the other two models.

Conclusions

The results reveal that the network model with a depth of 5 exhibits superior performance, albeit at the expense of the longest training time and maximum computational memory in the three models. A small server with two NVIDIA GeForce RTX 3090 GPUs with 24 G of memory was employed for this training. For the 3D U-Net model with a depth of more than 5 cannot be supported due to insufficient training memory, the 3D U-Net neural network with a depth of 5 is the commonly used and optimal choice for small servers.

Application and progress of artificial intelligence in radiation therapy dose prediction

Radiation therapy (RT) nowadays is a main treatment modality of cancer. To ensure the therapeutic efficacy of patients, accurate dose distribution is often required, which is a time-consuming and labor-intensive process. In addition, due to the differences in knowledge and experience among participants and diverse institutions, the predicted dose are often inconsistent. In last several decades, artificial intelligence (AI) has been applied in various aspects of RT, several products have been implemented in clinical practice and confirmed superiority. In this paper, we will review the research of AI in dose prediction, focusing on the progress in deep learning (DL).

PRT-Net: a progressive refinement transformer for dose prediction to guide ovarian transposition

Introduction

Young cervical cancer patients who require ovarian transposition usually have their ovaries moved away from the pelvic radiotherapy (RT) field before radiotherapy. The dose of ovaries during radiotherapy is closely related to the location of the ovaries. To protect ovarian function and avoid ovarian dose exceeding the limits, a safe location of transposed ovary must be determined prior to surgery.

Methods

For this purpose, we input the patient's preoperative CT into a neural network model to predict the dose distribution. Surgeons were able to quickly locate low-dose regions based on the dose distribution before surgery, thus determining the safe location of the transposed ovary. In this work, we proposed a new progressive refinement transformer model PRT-Net that can generate dose prediction at multiple scale resolutions in one forward propagation, and refine the dose prediction using prediction details from low to high resolution based on a deep supervision strategy. A multi-loss function fusion algorithm was also built to fit the prediction results under different loss dimensions. The clinical feasibility of the method was verified through an actual cases.

Results and discussion

Therefore, using PRT-Net to predict the dose distribution by preoperative CT in cervical cancer patients can assist clinicians to perform ovarian transposition surgery and prevent patients' ovaries from exceeding the prescribed dose limit in postoperative radiotherapy.

Novel in-house knowledge-based automated planning system for lung cancer treated with intensity-modulated radiotherapy

PURPOSE

The goal of this study was to propose a knowledge-based planning system which could automatically design plans for lung cancer patients treated with intensity-modulated radiotherapy (IMRT).

METHODS AND MATERIALS

From May 2018 to June 2020, 612 IMRT treatment plans of lung cancer patients were retrospectively selected to construct a planning database. Knowledge-based planning (KBP) architecture named αDiar was proposed in this study. It consisted of two parts separated by a firewall. One was the in-hospital workstation, and the other was the search engine in the cloud. Based on our previous study, A‑Net in the in-hospital workstation was used to generate predicted virtual dose images. A search engine including a three-dimensional convolutional neural network (3D CNN) was constructed to derive the feature vectors of dose images. By comparing the similarity of the features between virtual dose images and the clinical dose images in the database, the most similar feature was found. The optimization parameters (OPs) of the treatment plan corresponding to the most similar feature were assigned to the new plan, and the design of a new treatment plan was automatically completed. After αDiar was developed, we performed two studies. The first retrospective study was conducted to validate whether this architecture was qualified for clinical practice and involved 96 patients. The second comparative study was performed to investigate whether αDiar could assist dosimetrists in improving the quality of planning for the patients. Two dosimetrists were involved and designed plans for only one trial with and without αDiar; 26 patients were involved in this study.

RESULTS

The first study showed that about 54% (52/96) of the automatically generated plans would achieve the dosimetric constraints of the Radiation Therapy Oncology Group (RTOG) and about 93% (89/96) of the automatically generated plans would achieve the dosimetric constraints of the National Comprehensive Cancer Network (NCCN). The second study showed that the quality of treatment planning designed by junior dosimetrists was improved with the help of αDiar.

CONCLUSIONS

Our results showed that αDiar was an effective tool to improve planning quality. Over half of the patients' plans could be designed automatically. For the remaining patients, although the automatically designed plans did not fully meet the clinical requirements, their quality was also better than that of manual plans.

Patient-specific three-dimensional dose distribution prediction via deep learning for prostate cancer therapy: Improvement with the structure loss

PURPOSE

Deep learning (DL)-based dose distribution prediction can potentially reduce the cost of inverse planning process. We developed and introduced a structure-focused loss (L_struct) for 3D dose prediction to improve prediction accuracy. This study investigated the influence of L_struct on DL-based dose prediction for patients with prostate cancer. The proposed L_struct, which is similar in concept to dose-volume histogram (DVH)-based optimization in clinical practice, has the potential to provide more interpretable and accurate DL-based optimization.

METHODS

This study involved 104 patients who underwent prostate radiotherapy. We used 3D U-Net-based architecture to predict dose distributions from computed tomography and contours of the planning target volume and organs-at-risk. We trained two models using different loss functions: L2 loss and L_struct. Predicted doses were compared in terms of dose-volume parameters and the Dice similarity coefficient of isodose volume.

RESULTS

DVH analysis showed that the L_struct model had smaller errors from the ground truth than the L2 model. The L_struct model achieved more consistent dose distributions than the L2 model, with errors close to zero. The isodose Dice score of the L_struct model was greater than that of the L2 model by >20% of the prescribed dose.

CONCLUSIONS

We developed L_struct using labels of inputted contours for DL-based dose prediction for prostate radiotherapy. L_struct can be generalized to any DL architecture, thereby enhancing the dose prediction accuracy.

Automation of pencil beam scanning proton treatment planning for intracranial tumours

PURPOSE

To evaluate the feasibility of comprehensive automation of an intra-cranial proton treatment planning.

MATERIALS AND METHODS

Class solution (CS) beam configuration selection allows the user to identify predefined beam configuration based on target localization; automatic CS (aCS) will then explore all the possible CS beam geometries. Ten patients, already used for the evaluation of the automatic selection of the beam configuration, have been also employed to training an algorithm based on the computation of a benchmark dose exploit automatic general planning solution (GPS) optimization with a wish list approach for the planning optimization. An independent cohort of ten patients has been then used for the evaluation step between the clinical and the GPS plan in terms of dosimetric quality of plans and the time needed to generate a plan.

RESULTS

The definition of a beam configuration requires on average 22 min (range 9-29 min). The average time for GPS plan generation is 18 min (range 7-26 min). Median dose differences (GPS-Manual) for each OAR constraints are: brainstem -1.60 Gy, left cochlea -1.22 Gy, right cochlea -1.42 Gy, left eye 0.55 Gy, right eye -2.33 Gy, optic chiasm -1.87 Gy, left optic nerve -4.45 Gy, right optic nerve -2.48 Gy and optic tract -0.31 Gy. Dosimetric CS and aCS plan evaluation shows a slightly worsening of the OARs values except for the optic tract and optic chiasm for both CS and aCS, where better results have been observed.

CONCLUSION

This study has shown the feasibility and implementation of the automatic planning system for intracranial tumors. The method developed in this work is ready to be implemented in a clinical workflow.

Advances in Automated Treatment Planning

Treatment planning in radiation therapy has progressed enormously over the past several decades. Such advancements came in the form of innovative hardware and algorithms, giving rise to modalities such as intensity-modulated radiation therapy and volume modulated arc therapy, greatly improving patient outcome and quality of life. While these developments have improved the overall plan quality, they have also given rise to higher treatment planning complexity. This has resulted in increased treatment planning time and higher variability in the final approved plan quality. Radiation oncology, as an already technologically advanced field, has much research and implementation involving the use of AI. The field has begun to show the efficacy of using such technologies in many of its sub-areas, such as in diagnosis, imaging, segmentation, treatment planning, quality assurance, treatment delivery, and follow-up. Some AI technologies have already been clinically implemented by commercial systems. In this article, we will provide an overview to methods involved with treatment planning in radiation therapy. In particular, we will review the recent research and literature related to automation of the treatment planning process, leading to potentially higher efficiency and higher quality plans. We will then present the current and future challenges, as well as some future perspectives.

TransDose: a transformer-based UNet model for fast and accurate dose calculation for MR-LINACs

Objective. To present a transformer-based UNet model (TransDose) for fast and accurate dose calculation for magnetic resonance-linear accelerators (MR-LINACs). Approach. A 2D fluence map from each beam was first projected into a 3D fluence volume and then fed into the TransDose model together with patient density volume and output predicted beam dose. The proposed TransDose model combined a 3D residual UNet with a transformer encoder, where convolutional layers extracted the volumetric spatial features, and the transformer encoder processed the long-range dependencies in a global space. Ninety-eight cases with four tumor sites (brain, nasopharynx, lung, and rectum) treated with fixed-beam intensity-modulated radiotherapy were included in the dataset; 78 cases were used for model training and validation; and 20 cases were used for testing. The ground-truth beam doses were calculated with Monte Carlo (MC) simulations within 1% statistical uncertainty and magnetic field strength B = 1.5 T in the superior and inferior direction. Beam angles from the training and validation datasets were rotated 2–5 times, and doses were recalculated to augment the datasets. Results. The dose-volume histograms and indices between the predicted and MC doses showed good consistency. The average 3D γ-passing rates (3%/2 mm, for dose regions above 10% of maximum dose) were 99.13 ± 0.89% (brain), 98.31 ± 1.92% (nasopharynx), 98.74 ± 0.70% (lung), and 99.28 ± 0.25% (rectum). The average dose calculation time, which included the fluence projection and model prediction, was less than 310 ms for each beam. Significance. We successfully developed a transformer-based UNet dose calculation model—TransDose in magnetic fields. Its accuracy and efficiency indicated its potential for use in online adaptive plan optimization for MR-LINACs.

A Survey on Deep Learning for Precision Oncology

Precision oncology, which ensures optimized cancer treatment tailored to the unique biology of a patient’s disease, has rapidly developed and is of great clinical importance. Deep learning has become the main method for precision oncology. This paper summarizes the recent deep-learning approaches relevant to precision oncology and reviews over 150 articles within the last six years. First, we survey the deep-learning approaches categorized by various precision oncology tasks, including the estimation of dose distribution for treatment planning, survival analysis and risk estimation after treatment, prediction of treatment response, and patient selection for treatment planning. Secondly, we provide an overview of the studies per anatomical area, including the brain, bladder, breast, bone, cervix, esophagus, gastric, head and neck, kidneys, liver, lung, pancreas, pelvis, prostate, and rectum. Finally, we highlight the challenges and discuss potential solutions for future research directions.

Dose Prediction Using a Three-Dimensional Convolutional Neural Network for Nasopharyngeal Carcinoma With Tomotherapy

Purpose

This study focused on predicting 3D dose distribution at high precision and generated the prediction methods for nasopharyngeal carcinoma patients (NPC) treated with Tomotherapy based on the patient-specific gap between organs at risk (OARs) and planning target volumes (PTVs).

Methods

A convolutional neural network (CNN) is trained using the CT and contour masks as the input and dose distributions as output. The CNN is based on the "3D Dense-U-Net", which combines the U-Net and the Dense-Net. To evaluate the model, we retrospectively used 124 NPC patients treated with Tomotherapy, in which 96 and 28 patients were randomly split and used for model training and test, respectively. We performed comparison studies using different training matrix shapes and dimensions for the CNN models, i.e., 128 ×128 ×48 (for Model I), 128 ×128 ×16 (for Model II), and 2D Dense U-Net (for Model III). The performance of these models was quantitatively evaluated using clinically relevant metrics and statistical analysis.

Results

We found a more considerable height of the training patch size yields a better model outcome. The study calculated the corresponding errors by comparing the predicted dose with the ground truth. The mean deviations from the mean and maximum doses of PTVs and OARs were 2.42 and 2.93%. Error for the maximum dose of right optic nerves in Model I was 4.87 ± 6.88%, compared with 7.9 ± 6.8% in Model II (p=0.08) and 13.85 ± 10.97% in Model III (p<0.01); the Model I performed the best. The gamma passing rates of PTV₆₀ for 3%/3 mm criteria was 83.6 ± 5.2% in Model I, compared with 75.9 ± 5.5% in Model II (p<0.001) and 77.2 ± 7.3% in Model III (p<0.01); the Model I also gave the best outcome. The prediction error of D₉₅ for PTV₆₀ was 0.64 ± 0.68% in Model I, compared with 2.04 ± 1.38% in Model II (p<0.01) and 1.05 ± 0.96% in Model III (p=0.01); the Model I was also the best one.

Conclusions

It is significant to train the dose prediction model by exploiting deep-learning techniques with various clinical logic concepts. Increasing the height (Y direction) of training patch size can improve the dose prediction accuracy of tiny OARs and the whole body. Our dose prediction network model provides a clinically acceptable result and a training strategy for a dose prediction model. It should be helpful to build automatic Tomotherapy planning.