Methods, systems, and computer readable media for evaluating contours of physician-defined treatment volumes for radiotherapy
A convolutional neural network-based model predicts and highlights disagreements in physician-defined radiotherapy treatment volumes, improving segmentation accuracy and efficiency by providing quick feedback on treatment volume contours.
Patent Information
- Application Number
- PCT/US2026/026981
- Authority / Receiving Office
- WO · WO
- Patent Type
- Applications
- Current Assignee / Owner
- Priority Date
- 2025-05-07
- Filing Date
- 2026-05-06
- Publication Date
- 2026-11-12
AI Technical Summary
The segmentation of treatment volumes in radiotherapy, particularly for clinical target volumes and lymph nodes, is challenging due to the lack of well-defined boundaries and the reliance on clinical knowledge, which can be time-consuming and subjective, especially in settings without peer review.
A radiotherapy treatment volume segmentations/contours prediction model using an artificial convolutional neural network is trained on patient CT scans to predict and highlight regions of disagreement with physician-defined contours, facilitating peer review and improving segmentation accuracy.
The model provides quick voxel-wise segmentation predictions, enhancing the efficiency and accuracy of treatment volume definition by highlighting areas of agreement and disagreement, thus aiding in refining physician-defined contours.
Smart Images

Figure US2026026981_12112026_PF_FP_ABST
Abstract
Description
[0001] METHODS, SYSTEMS, AND COMPUTER READABLE MEDIA FOR EVALUATING CONTOURS OF PHYSICIAN-DEFINED TREATMENT VOLUMES FOR RADIOTHERAPY
[0002] GOVERNMENT INTEREST
[0003] This invention was made with government support under Grant No.
[0004] 1 R18HS029474 awarded by The Agency for Healthcare Research and Quality (AHRQ). The government has certain rights in the invention.
[0005] CROSS-REFERENCE TO RELATED APPLICATIONS This application claims the priority benefit of U.S. Provisional Patent Application Serial No. 63 / 801 ,775, filed May 7, 2023, the disclosure of which is incorporated herein by reference in its entirety.
[0006] TECHNICAL FIELD
[0007] The subject matter described herein relates to radiotherapy. More particularly, the subject matter described herein relates to methods, systems, and computer readable media for evaluating segmentation (e.g., contours) of physician-defined treatment volumes for radiotherapy, and for creating these segmentations (e.g., contours) to define treatment volumes.
[0008] BACKGROUND
[0009] A cornerstone of modern radiation therapy is the segmentation (contouring) of patient images into targets and normal tissues. Segmentation is commonly performed on the treatment planning computed tomography (CT) scan acquired prior to treatment. The definition of target volumes is often challenging and time consuming, requiring years of experience and clinical knowledge. This is especially true in the case of clinical target volumes (CTVs) which include regions of likely microscopic disease spread. A vital part of quality assurance at many healthcare facilities involves the evaluation of these target structures by a team of radiation oncology professionals (e.g. physicians, physicists, and dosimetrists familiar with this work). This process of contour / sim review involves the target contours from one radiation oncologist being reviewed by peers / colleagues. This is enabled by viewing thetarget volume with a series of two-dimensional images within the treatment planning system. The treating physician then can adjust their segmentations / contours based on the feedback and expertise from their peers and colleagues.
[0010] In healthcare facilities that do not have access to a team of radiation oncologists, the segmentations / contours within the treatment volume may be defined by a single radiation oncologist without the benefit of peer review. In either case, the task of defining segmentations / contours of the treatment volume can be challenging, especially when the targets rely heavily on clinical knowledge and lack well-defined boundaries between the regions that should be treated and regions that should be avoided. For example, it can be relatively easy to define the boundaries of an organ, such as the bladder, that is often avoided in radiotherapy treatment of the prostate gland. However, it can be comparatively more difficult to define contours of lymph nodes that drain the prostate gland, which do not have well-defined boundaries in treatment planning images. In addition, even when a team of radiation oncologists participates in a peer review session, individuals may be hesitant to speak up and change the treatment contours defined by their colleagues.
[0011] Considering these other challenges, there exists a need for improved methods, systems, and computer readable media for evaluating segmentations / contours of physician-defined treatment volumes for radiotherapy. Furthermore, tools that can define these nodal areas to be targeted, even in the setting where there are not already physician-defined segmentations / contours to review.
[0012] SUMMARY
[0013] A method for predicting segmentations / contours of treatment volumes for radiotherapy includes training a radiotherapy treatment volume segmentations / contours prediction model using an artificial convolutional neural network to predict segmentations / contours of the radiotherapy treatment volume for a subject. The model takes as an input the patient CT scan used for treatment planning. The model’s final output is a generated prediction target volume. This generated prediction volume is presented as a criticism to the physician’s defined target volume, with the regions ofdisagreement highlighted in green or red for under / over coverage, respectively. While the generation of Al predictions only require the patient CT of the patient in question, the intent of the model is to be used in conjunction with physician segmentation.
[0014] The subject matter described herein can be implemented in software in combination with hardware and / or firmware. For example, the subject matter described herein can be implemented in software executed by a processor. In one exemplary implementation, the subject matter described herein can be implemented using a non-transitory computer readable medium having stored thereon computer executable instructions that when executed by the processor of a computer control the computer to perform steps. Exemplary computer readable media suitable for implementing the subject matter described herein include non-transitory computer-readable media, such as disk memory devices, chip memory devices, programmable logic devices, and application specific integrated circuits. In addition, a computer readable medium that implements the subject matter described herein may be located on a single device or computing platform or may be distributed across multiple devices or computing platforms.
[0015] BRIEF DESCRIPTION OF THE DRAWINGS
[0016] Exemplary implementations of the subject matter described herein will now be explained with reference to the accompanying drawings, of which:
[0017] Figures 1A-1C illustrate an example of the peer review process for a prostate cancer patient with elective nodal volumes;
[0018] Figures 2A and 2B illustrate CT image slices with segmented / contoured regions to be avoided and to be treated;
[0019] Figure 3 illustrates another example of a segmented / contoured region to be treated during prostate gland radiotherapy;
[0020] Figure 4 is a diagram illustrating an overview of a radiotherapy treatment volume segmentations / contours prediction model training process;
[0021] Figure 5 is a diagram illustrating an overview of how the radiotherapy treatment volume segmentations / contours prediction model processes, duringsimulation review, a new CT image with physician-defined treatment segmentations / contours;
[0022] Figure 6 is an example of a radiotherapy treatment planning image after being processed by the radiotherapy treatment volume segmentations / contours prediction model;
[0023] Figure 7 is a 2D histogram of clinical treatment volumes;
[0024] Figure 8 is a 3D histogram of voxels of different types present in CT images for different patients;
[0025] Figure 9 illustrates normalized images of a pelvic nodal region of a patient;
[0026] Figure 10 is a diagram illustrating a structure of an attention UNet used to implement the radiotherapy treatment volume segmentations / contours prediction model;
[0027] Figure 11 is a graph illustrating evaluation of different combinations of hyperparameters for the radiotherapy treatment volume segmentations / contours prediction model;
[0028] Figure 12 is a block diagram of a deployed radiotherapy treatment volume contours prediction model that processes CTV images to evaluate contours of treatment volumes in DICOM formatted images;
[0029] Figure 13 is a flow chart illustrating an exemplary process for predicting contours of a radiotherapy treatment volume; and
[0030] Figure 14 illustrates example regions of disagreement most commonly observed between model predictions and clinically defined contours. In the left-hand images, disagreement relates to inclusion of the obturator lymph nodes. In the right-hand images, disagreement relates to the anterior aspect of the nodal target when initiating segmentation.
[0031] DETAILED DESCRIPTION
[0032] In radiotherapy, peer review (sometimes termed simulation review or contour review) is the process in which radiation oncologists review treatment planning images and associated segmentations / contours of a radiotherapytreatment target (and avoidance) volumes for a specific patient. Peer / simulation review can be subjective and defining radiotherapy contours can be a difficult, time-consuming task, especially for elective lymph nodes. The reason for this is that the lymph node chain does not present well-defined boundaries, relying on physician knowledge and experience.
[0033] Figures 1A-1C illustrate an example of the peer / simulation review process for a prostate cancer patient with elective nodal volumes. Figure 1 illustrates a team of radiation oncologists participating in a peer / simulation review. Figure 1 B (top image) is a CT image slice of a patient viewed from the perspective illustrated by the avatar in the lower image in Figure 1 B. The boundary adjacent to the pointer in the top image surrounds the bladder. The bladder is well defined in the CT image and is usually manually contoured by a radiation oncologist as an region to be avoided during radiotherapy treatment of the patient’s prostate gland and surrounding lymph nodes. In the top image in Figure 1 C, the bladder and the patient’s large intestines are segmented / contoured as regions to be avoided during radiotherapy. The right-most images in Figure 1 C show an avatar that illustrates the perspective of the left-most images in Figure 1C.
[0034] Figures 2A and 2B illustrate CT image slices with contoured regions to be avoided and to be treated. In the CT image of Figure 2A, the bladder is contoured as a region to be avoided. In the CT image of Figure 2B, regions between the bladder, femoral heads, and bowels are contoured by radiation oncologists as areas requiring radiotherapy treatment.
[0035] Figure 3 illustrates another example of a contoured region to be treated during prostate gland radiotherapy. In Figure 3, the contoured region includes lymph node areas surrounding / draining the prostate gland that may include microscopic cancer cells that are not visible in the CT image. In other words, the targeted nodal areas are often not seen on the XCT images as discrete nodes, and the likely nodal-bearing areas are defined based on the normal anatomy.
[0036] Figure 4 is a diagram illustrating an overview of a radiotherapy treatment volume segmentations / contours prediction model training process. In Figure 4, the physicians (typically radiation oncologists) participating insimulation reviews are labeled A-D. The CT images show treatment volumes defined by the physicians during simulation reviews. The treatment volumes defined by the physicians are used as a combined radiotherapy treatment volume contours prediction model training dataset, which is labeled UNC. The CT images with the physician-defined treatment volumes are input as ground truth data to the radiotherapy treatment volume contours prediction model. During training of the model, CT image slices without physician-defined treatment volume contours are fed to the model, and the model is trained to predict contours of treatment volumes that are similar to those in the ground truth images. Hyperparameters used in the model training process are described in detail below.
[0037] Figure 5 is a diagram illustrating an overview of how the radiotherapy treatment volume contours prediction model processes, during simulation review, a new CT image with physician-defined treatment contours. In Figure 5, the computer terminal represents the defining of a treatment volume by radiation oncologists during simulation review. The images including the defined treatment volume are uploaded to a server and converted to DICOM format. The radiotherapy treatment volume contours prediction model evaluates the physician-defined contours and generates its own contours, which is illustrated in more detail in Figure 6. The Al-generated contours are formatted into treatment planning session (TPS) loadable objects and presented to the physicians during the simulation review so that the physicians can see the Al-generated contours, the physician-generated contour, and evaluate the differences between the contours. Evaluating the differences between the contours enables the physicians to determine whether and how to change the contours of the treatment volume (and the regions to be avoided) for radiotherapy treatment.
[0038] Figure 6 is an example of a treatment planning image after being processed by the radiotherapy treatment volume contours prediction model. In particular, Figure 6 illustrates one contour, which is the contour selected by a collection of radiation oncologists, and another contour predicted by the radiotherapy treatment volume contours prediction model described herein. The outer contour is the physician-defined contour, and the inner contour isthe Al-generated contour predicted by the model. The area between the two contours is an area that should be removed from the radiotherapy treatment volume. The contours may be shown in a specific color, such as red, to illustrate that the Al-generated contour does not agree with the physician-defined contour. Treatment planning contours that agree with each other may be shown in a different color, such as green, from the color that indicates disagreement between the contours.
[0039] The software and technologies described herein were created with the goal of providing quick (within several minutes) voxel-wise segmentation predictions on medical imaging. The creation of a segmentation model can be largely broken up into three steps: 1) Data identification, acquisition, curation, and preparation, 2) Model creation, training, optimization, and implementation, and 3) Model deployment. Overcoming challenges on the path to this goal required the development and adaptation of new and existing software.
[0040] 1 . Data Preparation
[0041] 1 .1 . Data Identification:
[0042] The first step in creating a machine learning model is the acquisition of necessary training data. Within the University of North Carolina at Chapel Hill workflow, the radiation therapy treatment planning system RayStation is employed for all external-beam radiation therapy treatments with Elekta linear accelerators.
[0043] Previously created code has characterized the entirety of the RayStation treatment planning system into a personally crafted data architecture. This architecture, which is unrelated to RayStation, represents every aspect of information related to a patient's radiation treatment plan as represented by specific classes: Patient Class, Examination Class, Region of Interest, etc. Each patient is represented within a series of JavaScript object notation (JSON) files, enabling rapid identification of desired data without the use of other commercially available software.
[0044] The computer code for identifying patients first identifies treatment plans which were clinically approved for treatment. The computer code next searches for regions of interest (here-after referred to as ROIs) of type clinicaltarget volume (CTV) and planning target volume (PTV) and containing the name ‘CTV Pelvis’. This name is the clinical standard at our hospital to define this elective nodal volume.
[0045] The output from this step is a list of patients which will be of potential use. These include the database name, the medical record number (MRN), approving physician, the name of the ROI, the volume (in cc’s) of the ROI, the ROI Type (CTV, PTV), name of the associated case, and name of the examination.
[0046] 1 .2. Data Acquisition
[0047] Before the data is acquired from RayStation, a simple curation is made based on the volume of the identified ROI. Figure 7 is a histogram of values of CTV volumes, which helps identify potential outliers that may not represent the normal contours seen within clinical practice. From the CTV volumes illustrated in Figure 7, it was decided that any volume less than 200 cc or greater than 701 cc would be excluded. Software was then written to export the data for the identified patients from the treatment planning system in the form of DICOM files for the CT and .mh files for the defined clinical contour.
[0048] 1 .3. Data Conversion
[0049] The CT scans, presented in DICOM files, are not easily integrated into deep learning model training. Often data is converted into 3D NumPy arrays. For our workflow, we prefer to work with .nii files (hereafter referred to as neuroimaging informatics technology initiative (Nifti) files) as intermediaries. These file formats are easy to visualize with free software (vv-slicer) and evaluate the medical image overlapping with the ground truth segmentation.
[0050] Converting DICOM into Nifti files is not an especially easy task. However, software for converting from DICOM into Nifti format is described in Anderson et. al, “Simple Python Module for Conversions Between DICOM Images and Radiation Therapy Structures, Masks, and Prediction Arrays, Practical Radiation Oncology , Technical Report, Volume 11 , Issue 3, pp. 226-229 (May-June 2023) (https: / / doi.Org / 10.1016 / j.prro.2021.02.003), the disclosure of which is incorporated herein by reference in its entirety.1 .4. Data Moving
[0051] Manipulating large amounts of data on network drives is a slow process. Since the conversion of the data from DICOM into Nifti format removes all identifying information, the anonymized Nifti files are copied with an anonymized index to the computer locally. The excel file is updated to include a reference to the anonymized ID.
[0052] 1 .5. Data Conversion: Convert masks to Nifti Format
[0053] While the CT DICOM has been converted to Nifti, the mask contours remain as .mhd. Code has been designed to convert the masks from .mhd to the Nifti format, and add in a second-check of the presented volume, which is compared to that of the histogram distribution illustrated in Figure 7.
[0054] 1 .6. Data Conversion: Generate training data libraries
[0055] There are two main libraries for the creation of deep learning models: PyTorch and TensorFlow. PyTorch benefits by having a very relaxed form of data presentation and model creation; allowing if-else loops and variations to the model during the training process. TensorFlow has stricter requirements for model creation and training, largely maintaining abstract form that cannot vary during the training process.
[0056] Both libraries were investigated at different times, however, the final model creation has largely centered about TensorFlow. The creation of TensorFlow records (.tfrecords) allows us to create streamlined generators for our future models. This is essentially converting data into binary files that can be efficiently read and manipulated by our future models. After converting the files to binary format, the data is resampled into a uniform voxel size of 1.25 mm x 1.25 mm x 3.0 mm. The input image was also converted into a uniform 128 slices.
[0057] 1 .7. Data Evaluation
[0058] Normalizing the image values presented to the deep learning model is vital to facilitating the training process. In convolutional neural networks, theinitial kernels are random values with a mean of 0 and standard deviation of 1 . To maximize initial activation, the input image voxels should also be of value mean 0 and standard deviation of 1.
[0059] Figure 8 is a histogram of the voxels present within the image per patient. In Figure 8, the sample numbers represent image samples for individual patients. The bin values are voxel intensity values that fall within a particular statistical bin. The frequencies are counts of voxel intensities within each bin. From the data illustrated in Figure 8, we calculate a global mean and standard deviation which will be applied later during the training process. The histogram also helps identify meaningful ranges from which to ‘clip’ the input images. From the data illustrated in Figure 8, we identify a mean value of -35 Hounsfield Units (HU) and a standard deviation of 60. This value is used to normalize all input images, and clipped with an upper and lower bound of 5 standard deviations. Figure 9 illustrates an example of normalized images of a CTV for a patient.
[0060] 1 .8. Data Splitting
[0061] The final part of data preparation is distribution into training, testing, and validation. It is vital that data is separated by patient, as the similarity in anatomy from one scan to the next, if split into the training and testing group, would artificially indicate improved performance. Data was also separated based on the training physician. Each physician had data split into an 80 / 20 split, training to validation, where each patient must be grouped entirely within either the training or validation split.
[0062] 2. Model Creation
[0063] 2.1 . Data Generators
[0064] When creating a deep learning model, whether with PyTorch or TensorFlow, the method of providing example data to the model is handled via ‘generators’. These function by providing a stream of example data and often contain a series of pre-processing steps and perturbations. The perturbations help improve model stability by varying the input and providingmore examples to the model. The perturbations often include the injections of random noise, shifts, and random flipping of the images.
[0065] We created a series of unique processors for the data generators. First, images are normalized to a mean of 0 and standard deviation of 1 from the voxel intensity values described in section 1.7 above. Next, we created code to pull randomly cropped sections from the input data. This helps reduce the memory overhead of training on the entire 128x512x512 image and diversifies the training data throughout the training process. We investigated a variety of crop dimensions. Next, we applied a random flip in the left-right direction, as the volume which we are predicting has generally left-right symmetry. Lastly, we clip the image values to a range of -5 to +5.
[0066] 2.2. Deep learning models
[0067] Within the creation of a deep learning model we investigated fully convolutional networks. There are a number of hyperparameters which affect the overall structure of the model. We investigated creating ‘vanilla’ UNets, residual UNets, and the attention UNet. Figure 10 illustrates an example implementation of a radiotherapy treatment volume contours prediction model 1000, which includes a UNet architecture and the respective change in the concatenation layer which represents the attention net. More particularly, the left-hand side of Figure 10 illustrates the vanilla or plain UNet architecture, and the right-hand side of Figure 10 is the attention network. To form the attention UNet, the attention network on the right hand side of Figure 10 is connected where each of the green upsampling arrows are located in the UNet architecture on the left hand side of Figure 10.
[0068] Although in Figure 10, radiotherapy treatment volume contours prediction model 1000 is implemented using an attention UNet, the subject matter described herein is not limited to using an attention UNet to implement radiotherapy treatment volume contours prediction. Any suitable artificial neural network, including other convolutional neural networks, that can be trained to predict treatment volumes for radiotherapy from physician-created treatment volume contours is intended to be within the scope of the subject matter described herein.In Figure 10, radiotherapy treatment volume contours prediction model 1000 receives, as input, images that includes contours of a physician-defined radiotherapy treatment volume for a subject and generates, as output, images that include contours of a predicted treatment volume predicted by radiotherapy treatment volume contours prediction model 1000.
[0069] Radiotherapy treatment volume contours prediction model 1000 may be implemented on a computing platform 1002 including at least one processor 1004 and memory 1006. In one implementation, radiotherapy treatment volume contours prediction model 1000 may be implemented using computerexecutable instructions stored in memory 1006 and executed by processor 1004.
[0070] We investigated the number of layers to use, the number of convolution groups per layer, initial features, max features, value by which to increase or decrease the feature number, loss function, learning rate, batch size, and dropout. Hyper-parameters were evaluated using TensorBoard to assist in selecting the highest performing variables, a subsampling of which is shown in Figure 11 .
[0071] 2.3. Model Evaluation
[0072] The final radiotherapy treatment volume contours prediction model is evaluated based on a number of metrics: the Dice similarity coefficient, Jaccard, mean surface distance, and max surface distance. Previously created, publicly available, code is used to generate these performance metrics on the predicted images (https: / / github.com / brianmanderson / Segmentation_Evaluation_Tools).
[0073] 3. Model Deployment
[0074] Once a radiotherapy treatment volume contours prediction model has been created it still must be set up in a way which enables it to make predictions and provide feedback. This software serves as a prediction server for medical imaging data. It is composed of four primary components. Figure 12 is a block diagram illustrates a block diagram of a deployed prediction model that processes CTV images to evaluate contours of treatment volumesin DICOM formatted images. The components of the deployed model are as follows.
[0075] 3.1 . Listener
[0076] A dedicated listener continuously monitors a specified network drive for incoming DICOM files. Once a new DICOM series is detected and a ‘Completed’ text file is written, it triggers the remaining steps in the pipeline.
[0077] 3.2. Preprocessing
[0078] After detection, the DICOM data undergoes preprocessing, which includes converting the DICOM files into the format necessary for model prediction. Any necessary image normalization, resizing, or metadata handling occurs here.
[0079] 3.3. Postprocessing
[0080] The model output is then mapped back to the original image dimensions. This ensures that the predicted results align spatially with the original DICOM series. The resulting data is subsequently re-encoded into DICOM format as RT Structure files.
[0081] 3.4. Sender
[0082] Finally, the software generates a DICOM RT (Radiotherapy) structure containing the prediction output. This structure is transmitted back to the originating system (or designated network location), completing the prediction cycle. The sender includes radiotherapy treatment volume contours prediction model 1000 described above.
[0083] By organizing the pipeline into these four stages — monitoring, preprocessing, post-processing, and result delivery — the software provides an automated, end-to-end solution for running prediction models on incoming DICOM datasets and returning the results in a standardized medical imaging format. Furthermore, the modular structure of the server allows individual models to easily be added and removed without impacting other models.
[0084] Figure 13 is a flow chart illustrating an exemplary process for predicting contours of a radiotherapy treatment volume. Referring to Figure 13, in step 1300, the process includes training a radiotherapy treatment volume contours prediction model implemented using an artificial neural network to predict contours of a radiotherapy treatment volume for a subject. For example, amodel implemented using an artificial neural network, such as the attention UNet illustrated in Figure 10 can be trained by feeding the model with series of CT images with physician-generated radiotherapy treatment volume contours and CT images without physician-generated radiotherapy treatment volume contours and training the model to predict treatment volume contours that are like those in the input images with physician-defined contours. If the contours are generated by physicians in a particular healthcare system or institution, then the model will be trained to predict treatment volume contours that are like those manually generated by physicians in the healthcare system or institution.
[0085] In step 1302, the process further includes inputting, to the trained radiotherapy treatment volume contours prediction model, a series of images that include physician-defined contours of a radiotherapy treatment volume for a subject. For example, physicians may generate contours of a treatment volume on CT images of a subject using simulation review, and the images containing these contours may be input to the trained radiotherapy treatment volume contours prediction model.
[0086] In step 1304, the process further includes receiving, as output from the trained radiotherapy treatment volume contours prediction model, images that include contours of a predicted treatment volume predicted by the trained radiotherapy treatment volume contours prediction model and indicate agreement or disagreement with the physician-drawn contours. For example, the trained radiotherapy treatment volume contours prediction model may generate a series of images which include Al-generated contours of the treatment volume and physician generated contours of the treatment volume during the simulation review. Figure 6 illustrates one example of an image with Al-generated treatment volume contours.
[0087] In step 1306, the process further includes using the contours predicted by the trained radiotherapy treatment volume contours prediction model to plan radiotherapy treatment of the subject. For example, the Al-generated treatment volume contours may be used to refine the physician-generated treatment volume contours by moving the physician-defined treatment volumecontours so that the physician-defined treatment volume contours agree or nearly agree with the Al-generated treatment volume contours.
[0088] Discussion of Study Illustrating Training and Performance of Radiotherapy Treatment Contours Prediction Model This section describes a study in which the radiotherapy treatment volume contours prediction model was trained to predict radiotherapy treatment volumes, and the predicted volumes are compared to treatment volumes defined by experts.
[0089] Study Design and Participants
[0090] Two hundred twenty segmented treatment planning pelvic computed tomography scans from 216 patients irradiated for prostate cancer from 2021 to 2024 were analyzed (four patients having two scans due to weight loss and re-simulation). Each of these segmented scans had been through multidisciplinary peer review, including board-certified radiation oncologists with specialized expertise / experience in prostate cancer, physicists, and dosimetrists with experience in RT planning for patients with prostate cancer. These image segmentations, reflecting the collective input from our multidisciplinary peer review team, and used during the patients clinical planning, were taken as expert consensus. These patients typically received 180 cGy / fraction x 25 fractions = 4500 cGy or 180 cGy / fraction x 28 fractions = 5040 cGy to the nodal volumes, while simultaneously receiving 250 cGy / fraction x 28 fractions = 7000 cGy to the intact prostate gland. Patient images were split into an 80:20 training and internal validation set, resulting in 179 training and 41 validation images. External validation was performed on 10 completely withheld prostate nodal irradiation cases. Data were analyzed from March 2025 to June 2025.
[0091] Model Development
[0092] The DLS model development consisted of four stages as outlined below:Stage 1 : Curation of patient cohort. This broad search included all male patients who had received elective pelvic nodal irradiation. This cohort was further refined based on volume of the 23 contoured region named elective nodal CTV Pelvis and manually reviewed to identify the final 216 patients.
[0093] Stage 2: Image standardization and normalization was applied by resampling all images to a uniform resolution of 1.25 mm x 1.25 mm x 3.0 mm. Images were normalized to a range of -4 to 4 via subtraction of the mean, -35 HU, and division by the standard deviation, 60 HU. The resultant normalized image-centered values within the elective nodal CTV Pelvis are close to 0, beneficial for training newly created convolutional neural networks to focus on the desired region (see Figure 9 for an example).
[0094] Stage 3: Random perturbations were applied to enhance the overall training process. This included injections of random noise including intensity shifts with a maximum value of 1 .25 and randomly flipping images left / right to reduce overfitting.
[0095] Stage 4: Several deep learning segmentation model architectures were investigated; all centered about the UNet design. We varied the number of layers ‘deep’ between the encoding and decoding stages, the number of convolution blocks per layer, the number of channels in each block, and the addition of the ‘Attention Block’. Model selection was determined via categorical cross entropy loss. This model was then used to define AI / DLS-based image segmentations for electively treated nodal volumes.
[0096] Segmentation Performance
[0097] Segmentation performance (concordance with expert-defined contours) was measured using the Dice Similarity Coefficient (DSC), Median Surface Distance (MSD), and Hausdorff Distance (HD) to quantify overlap and spatial differences between DLS predicted segmentations using both internal and external validation datasets. All metrics were computed using Python version 3.11 with libraries such as SimplelTK and NumPy.RESULTS
[0098] Characteristics of the Dataset Summary of dataset characteristics are provided in T able 1 . Imaging parameters across datasets included a voxel size median (min-max) of 1.171 mm (0.9765-1.562 mm) x 1.171 mm (0.9765-1 .562 mm) x 2 mm (1 .5-3 mm).
[0099] Table 1 : Characteristics of the dataset
[0100] Training Internal External Characteristic Overall
[0101] Set Validation Validation Number of Image
[0102] 220 179 41 10
[0103] Sets
[0104] Number of Patients 216 175 41 10 Median Age, years 70 (se71 (52- 69 (51-86) 74 (62-90)
[0105]
[0106] (range) al) _ 91) _
[0107] Table 2: Segmentation Performance Metrics for Internal and External Validation Datasets
[0108] Dice Similarity Median Surface Hausdorff Dataset Coefficient (95% Distance, mm Distance, mm ci) (95% Cl) (95% Cl) Internal
[0109] Validation 0.79 (0.68-0.89) 1.68 (0.45-2.91) 35.4 (0-83.3) (N=41)
[0110] External
[0111] 27.14 (15.85- Validation 0.79 (0.71-0.86) 2.07 (0.83-3.31)
[0112] 38.08)
[0113]
[0114] (N=10)
[0115] DISCUSSION
[0116] Segmenting the pelvic nodal-bearing areas in patients with prostate cancer intended to receive elective nodal RT can be challenging. Indeed, several atlases have been created to assist physicians / planners with this task.
[0117] 15-17 However, atlas-guided imaging segmentation often requires significant time, with the user repeatedly looking back-and-forth between the planning and atlas images over a relatively large three-dimensional volume (i.e. including multiple anatomic areas on sequential axial images). The Al-based image segmentations described herein may provide a useful alternative means to “check” physician / planner-defined segmentations. The differences between our Al-based segmentations and our expert-based segmentationsare similar to those reported in studies quantifying the degree of interphysician variations in segmentations. For example, Hall et al noted a mean Dice coefficient of 0.66 (±0.18) and 0.63 (±0.13), in the definitive and postoperative delineation of the pelvic lymph node-bearing areas in prostate cancer, respectively based on a consensus discussion by 18 expert radiation oncologists, with further help of expert urologists.18 Hertogh et al., performed a workshop-based study involving 86 radiation oncologists focused on proposing an easily applicable, reproducible, and practice-validated contouring guideline for pelvic nodal CTV.19 The pre- vs. post-workshop agreement amongst different physicians, as quantified by the Dice coefficient, was 0.736 (±0.084) and 0.823 (±0.07), respectively. These Dice coefficients might be considered to be impressively high given the large number of participants across many institutions, while bearing in mind these participants were aware their segmentations were being scrutinized. Qualitative analysis conducted by Hall et al., 18 identified 4 areas of interobserver variability where experts struggled to reach an agreement, including the: (1) superior border of the common iliac nodes; (2) transition between the external iliac and inguinal nodes; (3) inclusion of the periprostatic nodes; and (4) inclusion of the perirectal nodes. Interestingly, during the validation workshops conducted by Hertogh et al.,19 the first 3 points also were deemed a matter of debate, the latter not being an issue. A fourth point of debate addressed during the workshop was the inferior limit of the presacral area. In our study, we noted similar locations of recurrent disagreement between the clinically defined contour and the Al-based contouring approach (see examples in Figure 14).
[0118] Figure 14 illustrates example regions of disagreement most commonly observed between model predictions 1400 and clinically defined contours 1402. In the left-hand images, disagreement relates to inclusion of the obturator lymph nodes. In the right-hand images, disagreement relates to the anterior aspect of the nodal target when initiating segmentation.
[0119] Limitations Our research presents several limitations. First, our Al-model was based on inputs from a limited number of physicians at a single institution; and in an environment where pre-planning peer review is routinely performed (i.e. where there might be resultant social pressure to conform tolocal institutional norms). This reality likely accounts for the relatively favorable Dice Similarity Coefficients and Median Surface Distances observed in our study. Thus, we recognize that our specific Al model might not align with practices in other institutions. Nevertheless, our approach could be repeated on a per-institutional (or even a per-physician) basis to generate tools that might assist in the image segmentation and review process. At the other extreme, one could generate Al models based on the segmentations from numerous centers / physicians that may be more broadly applicable. Second, since the nodal target volumes are large, the metrics used to compare the Al-based segmentations to the clinical segmentations might not adequately reflect meaningful differences. For example, clinically relevant differences in a few limited areas might be numerically diluted by the large portions of agreement between the segmented volumes and their surfaces. This is particularly true for the Dice Coefficients and the Median Surface Distances. This is why we also considered the Hausdorff Distance to highlight regions of potentially significant disagreement regardless of broad agreement elsewhere. Third, all of the segmented images used in our study were derived from patients whose elective nodal areas were planned to be treated at 180cGy / fraction for 25 fractions to a total dose of 4500cGy or for 28 fractions to a total dose of 5040cGy. The resultant Al-based models may not be broadly applicable to patients treated with alternative fractionations. Nevertheless, separate Al-based models could be trained on different patient subgroups based on fractionation schedules. Nonetheless, our study possesses several significant strengths. First, the creation of Al-based image segmentations of electively targeted node-bearing tissues may be able to facilitate a “standard approach” to peer-review (e.g. where areas of potential variation may be made readily apparent in a uniform manner). Second, the Al-tools we have developed are vendor-neutral, facilitating their adoption and integration into any clinic, irrespective of the treatment planning system employed.
[0120] CONCLUSION Overall, our findings suggest that Al-based image segmentations may be a reasonable tool to facilitate peer review discussions in clinical settings. Such tools might be particularly useful in investigations.Radiation oncology centers with a single physician make up -40% of the -2350 RT centers in the US20 and thus the number of patients who might be impacted by these tools is large.
[0121] The disclosure of each of the following references is incorporated herein by reference in its entirety.
[0122] References
[0123] 1. Marks LB, Jackson M, Xie L, et al. The challenge of maximizing safety in radiation oncology. Pract Radiat Oncol. 2011 ;1 (1):2-14. doi:10.1016 / j.prro.2010.10.001
[0124] 2. Meyer JL, Kavanagh BD, Purdy JA, Timmerman R, eds. IMRT, IGRT, SBRT - Advances in the Treatment Planning and Delivery of Radiotherapy. Vol 40. S. Karger AG; 2007. doi:10.1159 / isbn.978-3- 318-01415-0
[0125] 3. Lawton OAF, Michalski J, El-Naqa I, et al. Variation in the definition of clinical target volumes for pelvic nodal conformal radiation therapy for prostate cancer. Int J Radiat Oncol Biol Phys. 2009;74(2):377-382. doi : 10.1016 / j.ijrobp.2008.08.003
[0126] 4. Ezzell G, Chera B, Dicker A, et al. Common error pathways seen in the RO-ILS data that demonstrate opportunities for improving treatment safety. Pract Radiat Oncol. 2018;8(2):123-132. doi:10.1016 / j.prro.2017.10.007
[0127] 5. Kerns JR, Stingo F, Followill DS, Howell RM, Melancon A, Kry SF.
[0128] Treatment Planning System Calculation Errors Are Present in Most Imaging and Radiation Oncology Core- Houston Phantom Failures. Int J Radiat Oncol Biol Phys. 2017;98(5):1197-1203. doi:10.1016 / j.ijrobp.2017.03.049
[0129] 6. Segedin B, Petrie P. Uncertainties in target volume delineation in radiotherapy - are they relevant and what can we do about them? Radiol Oncol. 2016;50(3):254-262. doi:10.1515 / raon-2016-0023 7. Vinod SK, Min M, Jameson MG, Holloway LC. A review of interventions to reduce inter-observer variability in volume delineation in radiationoncology. J Med Imaging Radiat Oncol. 2016;60(3):393-406. doi : 10.1111 / 1754-9485.12462
[0130] Vijayakumar S, Duggar WN, Packianathan S, Morris B, Yang CC. Chasing Zero Harm in Radiation Oncology: Using Pre-treatment Peer Review. Front Oncol. 2019;9:302. doi:10.3389 / fonc.2019.00302 Marks LB, Adams RD, Pawlicki T, et al. Enhancing the role of case-oriented peer review to improve quality and safety in radiation oncology: Executive summary. Pract Radiat Oncol. 2013;3(3):149-156. doi:10.1016 / j.prro.2012.11.010
[0131] Chera BS, Potters L, Marks LB. Restructuring our approach to peer review: A critical need to improve the quality and safety of radiation therapy. Pract Radiat Oncol. 2020;10(5):321-323. doi : 10.1016 / j.prro.2020.07.001
[0132] Brunskill K, Nguyen TK, Boldt RG, et al. Does peer review of radiation plans affect clinical care? A systematic review of the literature. Int J Radiat Oncol Biol Phys. 2017;97(1):27-34. doi : 10.1016 / j . ijrobp.2016.09.015
[0133] Hoopes DJ, Johnstone PA, Chapin PS, et al. Practice patterns for peer review in radiation oncology. Pract Radiat Oncol. 2015;5(1 ):32-38. doi:10.1016 / j. prro.2014.04.004
[0134] Walburn T, Wang K, Sud S, et al. A prospective analysis of radiation oncologist compliance with early peer review recommendations. Int J Radiat Oncol Biol Phys. 2019;104(3):494-500. doi : 10.1016 / j .ijrobp .2019.02.036
[0135] Shumway JW, Adapa K, Amos A, et al. The Utility of Video-Based PreTreatment Peer Review in the COVID Era. Int J Radiat Oncol Biol Phys.
[0136] 2021 ;111 (3S):S94-S95. doi : 10.1016 / j. ijrobp.2021 .07.222
[0137] Harris VA, Staffurth J, Naismith O, et al. Consensus guidelines and contouring atlas for pelvic node delineation in prostate and pelvic node intensity modulated radiation therapy. Int J Radiat Oncol Biol Phys.
[0138] 2015;92(4):874-883. doi : 10.1016 / j . ijrobp.2015.03.021
[0139] Lawton CAF, Michalski J, El-Naqa I, et al. RTOG GU Radiation oncology specialists reach consensus on pelvic lymph node volumesfor high-risk prostate cancer. Int J Radiat Oncol Biol Phys.
[0140] 2009;74(2):383-387. doi : 10.1016 / j.ijrobp.2008.08.002
[0141] 17. Sargos P, Guerif S, Latorzeff I, et al. Definition of lymph node areas for radiotherapy of prostate cancer: A critical literature review by the French Genito-Urinary Group and the French Association of Urology (GETUG-AFU). Cancer Treat Rev. 2015;41 (10):814-820. doi:10.1016 / j.ctrv.2O15.10.005
[0142] 18. Hall WA, Paulson E, Davis BJ, et al. NRG oncology updated international consensus atlas on pelvic lymph node volumes for intact and postoperative prostate cancer. Int J Radiat Oncol Biol Phys.
[0143] 2021 ;109(1 ):174-185. doi:10.1016 / j.ijrobp.2020.08.034
[0144] 19. De Hertogh O, Le Bihan G, Zilli T, et al. Consensus delineation guidelines for pelvic lymph node radiation therapy of prostate cancer: on behalf of the francophone group of urological radiation therapy (GFRU). Int J Radiat Oncol Biol Phys. 2024;118(1):29-40. doi : 10.1016 / j.ijrobp.2023.07.020
[0145] 20. Milligan M, Hansen M, Kim DW, Orav EJ, Figueroa JF, Lam MB.
[0146] Practice consolidation among U.S. radiation oncologists over time. Int J Radiat Oncol Biol Phys. 2021 ;111 (3):610-618. doi : 10.1016 / j .ijrobp .2021 .06.009
[0147] It will be understood that various details of the subject matter described herein may be changed without departing from the scope of the subject matter described herein. Furthermore, the foregoing description is for the purpose of illustration only, and not for the purpose of limitation, as the subject matter described herein is defined by the claims as set forth hereinafter.
Claims
CLAIMSWhat is claimed is:1 . A method for predicting contours of treatment volumes for radiotherapy, the method comprising:training a radiotherapy treatment volume contours prediction model implemented using an artificial neural network to predict contours of a radiotherapy treatment volume for a subject;inputting, to the trained radiotherapy treatment volume contours prediction model, a series of images that include physician-defined contours of a radiotherapy treatment volume for a subject;receiving, as output from the trained radiotherapy treatment volume contours prediction model, images that include contours of a predicted treatment volume predicted by the trained radiotherapy treatment volume contours prediction model and indicate agreement or disagreement with the physician-drawn contours; andusing the contours predicted by the trained radiotherapy treatment volume contours prediction model to plan radiotherapy treatment of the subject.
2. The method of claim 1 , wherein training the radiotherapy treatment volume contours prediction model includes training the model using computed tomography (CT) images including radiotherapy treatment volume contours generated by physicians during simulation reviews.
3. The method of claim 2, wherein training the radiotherapy treatment volume contours prediction model includes adding perturbations to the CT images including radiotherapy treatment volume contours generated by physicians during simulation reviews.
4. The method of claim 3, wherein adding the perturbations to the CT images includes adding noise to the CT images, spatially shifting the CT images, and rotating the CT images.
5. The method of claim 1, wherein inputting the series of images to the trained radiotherapy treatment volume contours prediction model includes inputting a series of computed tomography (CT) images thatinclude contours of a treatment volume generated by physicians during simulation review.
6. The method of claim 1 , wherein receiving, as output, the images that include contours of the predicted treatment volume includes receiving a series of computed tomography (CT) images including artificial intelligence (Al)-predicted contours of the radiotherapy treatment volume and the physician-defined contours of the radiotherapy treatment volume.
7. The method of claim 1 , wherein receiving, as output, the images that include the contours of the treatment volume predicted by the model includes receiving the images during a simulation review to plan radiotherapy treatment for the subject.
8. The method of claim 1 , wherein the subject comprises a human subject seeking radiotherapy treatment for prostate cancer and the series of images including the physician-defined radiotherapy treatment volume contours includes images of a pelvic nodal region of the subject.
9. The method of claim 1 , wherein the artificial neural network comprises a convolutional neural network.
10. The method of claim 9, wherein the convolutional neural network comprises an attention UNet having a UNet architecture and an attention network connected between layers of the UNet architecture.
11. A system for predicting contours of treatment volumes for radiotherapy, the system comprising:a computing platform including at least one processor and a memory; anda trained radiotherapy treatment volume contours prediction model implemented using the at least one processor and comprising an artificial neural network trained to predict contours of a radiotherapy treatment volume for a subject, the trained radiotherapy treatment volume contours prediction model configured to:receive, as input, a series of images that include physician-defined contours of a radiotherapy treatment volume for a subject; andgenerate, as output, images that include contours of a predicted treatment volume predicted by the trained radiotherapy treatment volume contours prediction model and indicate agreement or disagreement with the physician-drawn contours.
12. The system of claim 11, wherein the trained radiotherapy treatment volume contours prediction model was trained using computed tomography (CT) images including radiotherapy treatment volume contours generated by physicians during simulation reviews.
13. The system of claim 12, wherein the CT images used in the training include spatial shifts and rotations.
14. The system of claim 11, wherein the series of images input to the trained radiotherapy treatment volume contours prediction model includes a series of computed tomography (CT) images that include contours of a treatment volume generated by physicians during simulation review.
15. The system of claim 11, wherein the images output from the radiotherapy treatment volume contours prediction model include a series of computed tomography (CT) images including artificial intelligence (Al)-predicted contours of the radiotherapy treatment volume and the physician-defined contours of the radiotherapy treatment volume.
16. The system of claim 11, wherein the radiotherapy treatment volume contours prediction model is configured to output the images during a simulation review to plan radiotherapy treatment for the subject.
17. The system of claim 11 , wherein the subject comprises a human subject seeking radiotherapy treatment for prostate cancer and the series of images including the physician-defined radiotherapy treatment volume contours includes images of a pelvic nodal region of the subject.
18. The system of claim 11 , wherein the artificial neural network comprises a convolutional neural network.
19. The system of claim 18, wherein the convolutional neural network comprises an attention UNet having a UNet architecture and an attention network connected between layers of the UNet architecture.
20. A non-transitory computer readable medium having stored thereon executable instructions that when executed by a processor of a computer control the computer to perform steps comprising:training a radiotherapy treatment volume contours prediction model implemented using an artificial neural network to predict contours of a radiotherapy treatment volume for a subject;inputting, to the trained radiotherapy treatment volume contours prediction model, a series of images that include physician-defined contours of a radiotherapy treatment volume for a subject;receiving, as output from the trained radiotherapy treatment volume contours prediction model, images that include contours of a predicted treatment volume predicted by the trained radiotherapy treatment volume contours prediction model and indicate agreement or disagreement with the physician-drawn contours; andusing the contours predicted by the trained radiotherapy treatment volume contours prediction model to plan radiotherapy treatment of the subject.
