Purpose Cone‐beam computed tomography (CBCT) scanning is used daily or weekly (i.e., on‐treatment CBCT) for accurate patient setup in image‐guided radiotherapy. However, inaccuracy of CT numbers prevents CBCT from performing advanced tasks such as dose calculation and treatment planning. Motivated by the promising performance of deep learning in medical imaging, we propose a deep U‐net‐based approach that synthesizes CT‐like images with accurate numbers from planning CT, while keeping the same anatomical structure as on‐treatment CBCT. Methods We formulated the CT synthesis problem under a deep learning framework, where a deep U‐net architecture was used to take advantage of the anatomical structure of on‐treatment CBCT and image intensity information of planning CT. U‐net was chosen because it exploits both global and local features in the image spatial domain, matching our task to suppress global scattering artifacts and local artifacts such as noise in CBCT. To train the synthetic CT generation U‐net (sCTU‐net), we include on‐treatment CBCT and initial planning CT of 37 patients (30 for training, seven for validation) as the input. Additional replanning CT images acquired on the same day as CBCT after deformable registration are utilized as the corresponding reference. To demonstrate the effectiveness of the proposed sCTU‐net, we use another seven independent patient cases (560 slices) for testing. Results We quantitatively compared the resulting synthetic CT (sCT) with the original CBCT image using deformed same‐day pCT images as reference. The averaged accuracy measured by mean absolute error (MAE) between sCT and reference CT (rCT) on testing data is 18.98 HU, while MAE between CBCT and rCT is 44.38 HU. Conclusions The proposed sCTU‐net can synthesize CT‐quality images with accurate CT numbers from on‐treatment CBCT and planning CT. This potentially enables advanced CBCT applications for adaptive treatment planning.

The dual-energy CT-based (DECT) approach holds promise in reducing the overall uncertainty in proton stopping-power-ratio (SPR) estimation as compared to the conventional stoichiometric calibration approach. The objective of this study was to analyze the factors contributing to uncertainty in SPR estimation using the DECT-based approach and to derive a comprehensive estimate of the range uncertainty associated with SPR estimation in treatment planning. Two state-of-the-art DECT-based methods, the Hünemohr-Saito method (2014, 2012) and the Bourque method (2014), were selected and implemented on a Siemens SOMATOM Force DECT scanner. The uncertainties were first divided into five independent categories. The uncertainty associated with each category was estimated for lung, soft and bone tissues separately. A single composite uncertainty estimate was eventually determined for three tumor sites (lung, prostate and head-and-neck) by weighting the relative proportion of each tissue group for that specific site. The uncertainties associated with the two selected DECT methods were found to be similar, therefore the following results applied to both methods. The overall uncertainty (1σ) in SPR estimation with the DECT-based approach was estimated to be 3.8%, 1.2% and 2.0% for lung, soft and bone tissues, respectively. The dominant factor contributing to uncertainty in the DECT approach was the imaging uncertainties, followed by the DECT modeling uncertainties. Our study showed that the DECT approach can reduce the overall range uncertainty to approximately 2.2% (2σ) in clinical scenarios, in contrast to the previously reported 1%.

PurposeIn the treatment planning process of intensity‐modulated radiation therapy (IMRT), a human planner operates the treatment planning system (TPS) to adjust treatment planning parameters, for example, dose volume histogram (DVH) constraints’ locations and weights, to achieve a satisfactory plan for each patient. This process is usually time‐consuming, and the plan quality depends on planer’s experience and available planning time. In this study, we proposed to model the behaviors of human planners in treatment planning by a deep reinforcement learning (DRL)‐based virtual treatment planner network (VTPN), such that it can operate the TPS in a human‐like manner for treatment planning.Methods and MaterialsUsing prostate cancer IMRT as an example, we established the VTPN using a deep neural network developed. We considered an in‐house optimization engine with a weighted quadratic objective function. Virtual treatment planner network was designed to observe an intermediate plan DVHs and decide the action to improve the plan by changing weights and threshold dose in the objective function. We trained the VTPN in an end‐to‐end DRL process in 10 patient cases. A plan score was used to measure plan quality. We demonstrated the feasibility and effectiveness of the trained VTPN in another 64 patient cases.ResultsVirtual treatment planner network was trained to spontaneously learn how to adjust treatment planning parameters to generate high‐quality treatment plans. In the 64 testing cases, with initialized parameters, quality score was 4.97 (±2.02), with 9.0 being the highest possible score. Using VTPN to perform treatment planning improved quality score to 8.44 (±0.48).ConclusionsTo our knowledge, this was the first time that intelligent treatment planning behaviors of human planner in external beam IMRT are autonomously encoded in an artificial intelligence system. The trained VTPN is capable of behaving in a human‐like way to produce high‐quality plans.

Inverse treatment planning in radiation therapy is formulated as solving optimization problems. The objective function and constraints consist of multiple terms designed for different clinical and practical considerations. Weighting factors of these terms are needed to define the optimization problem. While a treatment planning optimization engine can solve the optimization problem with given weights, adjusting the weights to yield a high-quality plan is typically performed by a human planner. Yet the weight-tuning task is labor intensive, time consuming, and it critically affects the final plan quality. An automatic weight-tuning approach is strongly desired. The procedure of weight adjustment to improve the plan quality is essentially a decision-making problem. Motivated by the tremendous success in deep learning for decision making with human-level intelligence, we propose a novel framework to adjust the weights in a human-like manner. This study uses inverse treatment planning in high-dose-rate brachytherapy (HDRBT) for cervical cancer as an example. We develop a weight-tuning policy network (WTPN) that observes dose volume histograms of a plan and outputs an action to adjust organ weighting factors, similar to the behaviors of a human planner. We train the WTPN via end-to-end deep reinforcement learning. Experience replay is performed with the epsilon greedy algorithm. After training is completed, we apply the trained WTPN to guide treatment planning of five testing patient cases. It is found that the trained WTPN successfully learns the treatment planning goals and is able to guide the weight tuning process. On average, the quality score of plans generated under the WTPN's guidance is improved by ~8.5% compared to the initial plan with arbitrarily set weights, and by 10.7% compared to the plans generated by human planners. To our knowledge, this is the first time that a tool is developed to adjust organ weights for the treatment planning optimization problem in a human-like fashion based on intelligence learnt from a training process. This is different from existing strategies based on pre-defined rules. The study demonstrates potential feasibility to develop intelligent treatment planning approaches via deep reinforcement learning.

As one of the most popular approaches in artificial intelligence, deep learning (DL) has attracted a lot of attention in the medical physics field over the past few years. The goals of this topical review article are twofold. First, we will provide an overview of the method to medical physics researchers interested in DL to help them start the endeavor. Second, we will give in-depth discussions on the DL technology to make researchers aware of its potential challenges and possible solutions. As such, we divide the article into two major parts. The first part introduces general concepts and principles of DL and summarizes major research resources, such as computational tools and databases. The second part discusses challenges faced by DL, present available methods to mitigate some of these challenges, as well as our recommendations.

A number of image-processing problems can be formulated as optimization problems. The objective function typically contains several terms specifically designed for different purposes. Parameters in front of these terms are used to control the relative importance among them. It is of critical importance to adjust these parameters, as quality of the solution depends on their values. Tuning parameters are a relatively straight forward task for a human, as one can intuitively determine the direction of parameter adjustment based on the solution quality. Yet manual parameter tuning is not only tedious in many cases, but also becomes impractical when a number of parameters exist in a problem. Aiming at solving this problem, this paper proposes an approach that employs deep reinforcement learning to train a system that can automatically adjust parameters in a human-like manner. We demonstrate our idea in an example problem of optimization-based iterative computed tomography (CT) reconstruction with a pixel-wise total-variation regularization term. We set up a parameter-tuning policy network (PTPN), which maps a CT image patch to an output that specifies the direction and amplitude by which the parameter at the patch center is adjusted. We train the PTPN via an end-to-end reinforcement learning procedure. We demonstrate that under the guidance of the trained PTPN, reconstructed CT images attain quality similar or better than those reconstructed with manually tuned parameters.

Purpose We propose a novel domain‐specific loss, which is a differentiable loss function based on the dose‐volume histogram (DVH), and combine it with an adversarial loss for the training of deep neural networks. In this study, we trained a neural network for generating Pareto optimal dose distributions, and evaluate the effects of the domain‐specific loss on the model performance. Methods In this study, three loss functions — mean squared error (MSE) loss, DVH loss, and adversarial (ADV) loss — were used to train and compare four instances of the neural network model: (a) MSE, (b) MSE + ADV, (c) MSE + DVH, and (d) MSE + DVH+ADV. The data for 70 prostate patients, including the planning target volume (PTV), and the organs at risk (OAR) were acquired as 96 × 96 × 24 dimension arrays at 5 mm3 voxel size. The dose influence arrays were calculated for 70 prostate patients, using a 7 equidistant coplanar beam setup. Using a scalarized multicriteria optimization for intensity‐modulated radiation therapy, 1200 Pareto surface plans per patient were generated by pseudo‐randomizing the PTV and OAR tradeoff weights. With 70 patients, the total number of plans generated was 84 000 plans. We divided the data into 54 training, 6 validation, and 10 testing patients. Each model was trained for a total of 100,000 iterations, with a batch size of 2. All models used the Adam optimizer, with a learning rate of 1 × 10−3. Results Training for 100 000 iterations took 1.5 days (MSE), 3.5 days (MSE+ADV), 2.3 days (MSE+DVH), and 3.8 days (MSE+DVH+ADV). After training, the prediction time of each model is 0.052 s. Quantitatively, the MSE+DVH+ADV model had the lowest prediction error of 0.038 (conformation), 0.026 (homogeneity), 0.298 (R50), 1.65% (D95), 2.14% (D98), and 2.43% (D99). The MSE model had the worst prediction error of 0.134 (conformation), 0.041 (homogeneity), 0.520 (R50), 3.91% (D95), 4.33% (D98), and 4.60% (D99). For both the mean dose PTV error and the max dose PTV, Body, Bladder and rectum error, the MSE+DVH+ADV outperformed all other models. Regardless of model, all predictions have an average mean and max dose error <2.8% and 4.2%, respectively. Conclusion The MSE+DVH+ADV model performed the best in these categories, illustrating the importance of both human and learned domain knowledge. Expert human domain‐specific knowledge can be the largest driver in the performance improvement, and adversarial learning can be used to further capture nuanced attributes in the data. The real‐time prediction capabilities allow for a physician to quickly navigate the tradeoff space for a patient, and produce a dose distribution as a tangible endpoint for the dosimetrist to use for planning. This is expected to considerably reduce the treatment planning time, allowing for clinicians to focus their efforts on the difficult and demanding cases.

Radiotherapy treatment planning currently requires many trial-anderror iterations between the planner and treatment planning system, as well as between the planner and physician for discussion/consultation. The physician's preferences for a particular patient cannot be easily quantified and precisely conveyed to the planner. In this study we present a real-time volumetric Pareto surface dose generation deep learning neural network that can be used after segmentation by the physician, adding a tangible and quantifiable endpoint to portray to the planner. From 70 prostate patients, we first generated 84,000 intensity modulated radiation therapy plans (1,200 plans per patient) sampling the Pareto surface, representing various tradeoffs between the planning target volume (PTV) and the organs-at-risk (OAR), including bladder, rectum, left femur, right femur, and body. We divided the data to 10 test patients and 60 training/validation patients. We then trained a hierarchically densely connected convolutional U-net (HD U-net), to take the PTV and avoidance map representing OARs masks and weights, and predict the optimized plan. The HD U-net is capable of accurately predicting the 3D Pareto optimal dose distributions, with average [mean, max] dose errors of [3.4%, 7.7%](PTV), [1.6%, 5.6%](bladder), [3.7%, 4.2%](rectum), [3.2%, 8.0%](left femur), [2.9%, 7.7%](right femur), and [0.04%, 5.4%](body) of the prescription dose. The PTV dose coverage prediction was also very similar, with errors of 1.3% (D98) and 2.0% (D99). Homogeneity was also similar, differing by 0.06 on average. The neural network can predict the dose within 1.7 seconds. Clinically, the optimization and dose calculation is much slower, taking 5-10 minutes.