Near Real-Time Assessment of Anatomic and Dosimetric Variations for Head and Neck Radiation Therapy via Graphics Processing Unit-based Dose Deformation Framework

Artificial general intelligence for radiation oncology

The emergence of artificial general intelligence (AGI) is transforming radiation oncology. As prominent vanguards of AGI, large language models (LLMs) such as GPT-4 and PaLM 2 can process extensive texts and large vision models (LVMs) such as the Segment Anything Model (SAM) can process extensive imaging data to enhance the efficiency and precision of radiation therapy. This paper explores full-spectrum applications of AGI across radiation oncology including initial consultation, simulation, treatment planning, treatment delivery, treatment verification, and patient follow-up. The fusion of vision data with LLMs also creates powerful multimodal models that elucidate nuanced clinical patterns. Together, AGI promises to catalyze a shift towards data-driven, personalized radiation therapy. However, these models should complement human expertise and care. This paper provides an overview of how AGI can transform radiation oncology to elevate the standard of patient care in radiation oncology, with the key insight being AGI's ability to exploit multimodal clinical data at scale.

A systematic review on clinical adaptive radiotherapy for head and neck cancer

INTRODUCTION

Head and neck cancer (HNC) patients' anatomy may undergo significant changes during radiotherapy (RT). This potentially affects dose distribution and compromises conformity between planned and delivered dose. Adaptive radiotherapy (ART) is a promising technique to overcome this problem but requires a significant workload. This systematic review aims to estimate the clinical and dosimetric benefits of ART using prospective data.

MATERIAL AND METHODS

A search on PubMed and Web of Science according to the PRISMA guidelines was made on Feb 6, 2023. Search string used was: 'adaptive radiotherapy head neck cancer'. English language filter was applied. All studies were screened for inclusion on title and abstract, and the full text was read and discussed in the research group in case of uncertainty. Inclusion criteria were a prospective ART strategy for HNC investigating clinical or dosimetric outcomes.

RESULTS

A total of 1251 articles were identified of which 15 met inclusion criteria. All included studies were published between 2010 and 2023 with a substantial diversity in design, endpoints, and nomenclature. The number of patients treated with ART was small with a median of 20 patients per study (range 4 to 86), undergoing 1-2 replannings. Mean dose to the parotid glands was reduced by 0.4-7.1 Gy. Maximum dose to the spinal cord was reduced by 0.5-4.6 Gy. Only five studies reported clinical outcome and disease control was excellent. Data on toxicity were ambiguous with some studies indicating reduced acute toxicity and xerostomia, while others found reduced quality of life in patients treated with ART.

CONCLUSION

The literature on clinical ART in HNC is limited. ART is associated with small reductions in doses to organs at risk, but the influence on toxicity and disease control is uncertain. There is a clear need for larger, prospective trials with a well-defined control group.

Deep-learning based fast and accurate 3D CT deformable image registration in lung cancer

BACKGROUND

Deformable Image Registration (DIR) is an essential technique required in many applications of radiation oncology. However, conventional DIR approaches typically take several minutes to register one pair of 3D CT images and the resulting deformable vector fields (DVFs) are only specific to the pair of images used, making it less appealing for clinical application.

PURPOSE

A deep-learning-based DIR method using CT images is proposed for lung cancer patients to address the common drawbacks of the conventional DIR approaches and in turn can accelerate the speed of related applications, such as contour propagation, dose deformation, adaptive radiotherapy (ART), etc. METHODS: A deep neural network based on VoxelMorph was developed to generate DVFs using CT images collected from 114 lung cancer patients. Two models were trained with the weighted mean absolute error (wMAE) loss and structural similarity index matrix (SSIM) loss (optional) (i.e., the MAE model and the M+S model). In total, 192 pairs of initial CT (iCT) and verification CT (vCT) were included as a training dataset and the other independent 10 pairs of CTs were included as a testing dataset. The vCTs usually were taken 2 weeks after the iCTs. The synthetic CTs (sCTs) were generated by warping the vCTs according to the DVFs generated by the pre-trained model. The image quality of the synthetic CTs was evaluated by measuring the similarity between the iCTs and the sCTs generated by the proposed methods and the conventional DIR approaches, respectively. Per-voxel absolute CT-number-difference volume histogram (CDVH) and MAE were used as the evaluation metrics. The time to generate the sCTs was also recorded and compared quantitatively. Contours were propagated using the derived DVFs and evaluated with SSIM. Forward dose calculations were done on the sCTs and the corresponding iCTs. Dose volume histograms (DVHs) were generated based on dose distributions on both iCTs and sCTs generated by two models, respectively. The clinically relevant DVH indices were derived for comparison. The resulted dose distributions were also compared using 3D Gamma analysis with thresholds of 3 mm/3%/10% and 2 mm/2%/10%, respectively.

RESULTS

The two models (wMAE and M+S) achieved a speed of 263.7±163 / 265.8±190 ms and a MAE of 13.15±3.8 / 17.52±5.8 HU for the testing dataset, respectively. The average SSIM scores of 0.987±0.006 and 0.988±0.004 were achieved by the two proposed models, respectively. For both models, CDVH of a typical patient showed that less than 5% of the voxels had a per-voxel absolute CT-number-difference larger than 55 HU. The dose distribution calculated based on a typical sCT showed differences of ≤2cGy[RBE] for clinical target volume (CTV) D95 and D5 , within ±0.06% for total lung V5 , ≤1.5cGy[RBE] for heart and esophagus Dmean , and ≤6cGy[RBE] for cord Dmax compared to the dose distribution calculated based on the iCT. The good average 3D Gamma passing rates (> 96% for 3 mm/3%/10% and > 94% for 2 mm/2%/10%, respectively) were also observed.

CONCLUSION

A deep neural network-based DIR approach was proposed and has been shown to be reasonably accurate and efficient to register the initial CTs and verification CTs in lung cancer.

Retrospective Clinical Evaluation of a Decision-Support Software for Adaptive Radiotherapy of Head and Neck Cancer Patients

Purpose

This study aimed to evaluate the clinical need for an automated decision-support software platform for adaptive radiation therapy (ART) of head and neck cancer (HNC) patients.

Methods

We tested RTapp (SegAna), a new ART software platform for deciding when a treatment replan is needed, to investigate a set of 27 HNC patients’ data retrospectively. For each fraction, the software estimated key components of ART such as daily dose distribution and cumulative doses received by targets and organs at risk (OARs) from daily 3D imaging in real-time. RTapp also included a prediction algorithm that analyzed dosimetric parameter (DP) trends against user-specified thresholds to proactively trigger adaptive re-planning up to four fractions ahead. The DPs evaluated for ART were based on treatment planning dose constraints. Warning (V₉₅<95%) and adaptation (V₉₅<93%) thresholds were set for PTVs, while OAR adaptation dosimetric endpoints of +10% (DE₁₀) were set for all D_max and D_mean DPs. Any threshold violation at end of treatment (EOT) triggered a review of the DP trends to determine the threshold-crossing fraction Fx when the violations occurred. The prediction model accuracy was determined as the difference between calculated and predicted DP values with 95% confidence intervals (CI₉₅).

Results

RTapp was able to address the needs of treatment adaptation. Specifically, we identified 18/27 studies (67%) for violating PTV coverage or parotid D_mean at EOT. Twelve PTVs had V₉₅<95% (mean coverage decrease of −6.8 ± 2.9%) including six flagged for adaptation at median Fx= 6 (range, 1–16). Seventeen parotids were flagged for exceeding D_mean dose constraints with a median increase of +2.60 Gy (range, 0.99–6.31 Gy) at EOT, including nine with DP>DE₁₀. The differences between predicted and calculated PTV V₉₅ and parotid D_mean was up to 7.6% (mean ± CI₉₅, −2.7 ± 4.1%) and 5 Gy (mean ± CI₉₅, 0.3 ± 1.6 Gy), respectively. The most accurate predictions were obtained closest to the threshold-crossing fraction. For parotids, the results showed that Fx ranged between fractions 1 and 23, with a lack of specific trend demonstrating that the need for treatment adaptation may be verified for every fraction.

Conclusion

Integrated in an ART clinical workflow, RTapp aids in predicting whether specific treatment would require adaptation up to four fractions ahead of time.

Evaluation of daily dose accumulation with deformable image registration method using helical tomotherapy images for nasopharyngeal carcinoma

Aim:

Nasopharyngeal carcinoma (NPC) patients may have anatomical variations during their radiotherapy treatment course. In this study, we determine the daily accumulated dose by the deformable image registration (DIR) process for comparing with the planned dose and explore the number of fractions which the daily accumulated dose significantly changed from the planned dose.

Methods:

The validation of the DIR process in MIM software has been tested. One hundred and sixty-five daily megavoltage computed tomography (MVCT) images of NPC patients who were treated by helical tomotherapy were exported to MIM software to determine the daily accumulated dose and then compared with the planned dose.

Results:

The MIM software illustrated the acceptable validation for clinical application. The accumulated dose (D50%) of the planning target volume (PTV70) showed a decrease from the planned dose with an average of 0.5 ± 0.27% at the end of the treatment and was significantly different from the planned dose after the second fraction of the treatment (p-value = 0.008). In contrast, the accumulated dose of organ at risk (OAR) tended to increase from the planned dose and was significantly different after the fifth fraction (left parotid), the twelfth fraction (right parotid) and the second fraction (spinal cord).

Findings:

The inter-fractional anatomic changes cause the actual dose to be different from the planned dose. The dose differences and the number of fractions were varied in each target and OAR. The dose accumulation explored the necessary information for the radiation oncologist to consider adaptive treatment strategies to increase the efficiency of treatment.

Single-Institute Clinical Experiences Using Whole-Field Simultaneous Integrated Boost (SIB) Intensity-Modulated Radiotherapy (IMRT) and Sequential IMRT in Postoperative Patients With Oral Cavity Cancer (OCC)

This study aimed to review clinical experiences using whole-field simultaneous integrated boost (SIB) intensity-modulated radiotherapy (IMRT) and sequential IMRT in postoperative patients with oral cavity cancer (OCC). From November 2006 to December 2014, a total of 182 postoperative patients with OCC who underwent either SIB-IMRT (n = 63) or sequential IMRT (n = 119) were enrolled retrospectively and matched randomly according to multiple risk factors by a computer. The differences were well balanced after patient matching (P = .38). The median follow-up time was 65 months. For patients treated with the SIB technique and the sequential technique, the respective mortality rates were 36.8% and 20.0% (P = .04). The primary recurrence rates were 26.3% and 10.0% (P = .02), respectively. The respective marginal failure rates were 26.7% and 16.7%. A multivariate logistic regression analysis showed that patients who received the SIB technique had a 2.74 times higher risk of death than those who received the sequential technique (95% confidence interval = 1.10-6.79, P = .03). Sequential IMRT provided a significantly lower dose to the esophagus (5.2 Gy, P = .02) and trachea (4.6 Gy, P = .03) than SIB-IMRT. For patients with locally advanced OCC, postoperative sequential IMRT may overcome an unpredictable geographic miss, potentially with a lower marginal failure rate in the primary area. Patients treated by sequential IMRT show equal overall survival benefits to those treated by SIB-IMRT and a lower mortality rate than those treated by SIB-IMRT. Additionally, a reduced dose to the esophagus and trachea compared to sequential IMRT was noted.

An adversarial machine learning framework and biomechanical model-guided approach for computing 3D lung tissue elasticity from end-expiration 3DCT

PURPOSE

Lung elastography aims at measuring the lung parenchymal tissue elasticity for applications ranging from diagnostic purposes to biomechanically guided deformations. Characterizing the lung tissue elasticity requires four-dimensional (4D) lung motion as an input, which is currently estimated by deformably registering 4D computed tomography (4DCT) datasets. Since 4DCT imaging is widely used only in a radiotherapy treatment setup, there is a need to predict the elasticity distribution in the absence of 4D imaging for applications within and outside of radiotherapy domain.

METHODS

In this paper, we present a machine learning-based method that predicts the three-dimensional (3D) lung tissue elasticity distribution for a given end-expiration 3DCT. The method to predict the lung tissue elasticity from an end-expiration 3DCT employed a deep neural network that predicts the tissue elasticity for the given CT dataset. For training and validation purposes, we employed five-dimensional CT (5DCT) datasets and a finite element biomechanical lung model. The 5DCT model was first used to generate end-expiration lung geometry, which was taken as the source lung geometry for biomechanical modeling. The deformation vector field pointing from end expiration to end inhalation was computed from the 5DCT model and taken as input in order to solve for the lung tissue elasticity. An inverse elasticity estimation process was employed, where we iteratively solved for the lung elasticity distribution until the model reproduced the ground-truth deformation vector field. The machine learning process uses a specific type of learning process, namely a constrained generalized adversarial neural network (cGAN) that learned the lung tissue elasticity in a supervised manner. The biomechanically estimated tissue elasticity together with the end-exhalation CT was the input for the supervised learning. The trained cGAN generated the elasticity from a given breath-hold CT image. The elasticity estimated was validated in two approaches. In the first approach, a L2-norm-based direct comparison was employed between the estimated elasticity and the ground-truth elasticity. In the second approach, we generated a synthetic four-dimensional CT (4DCT0 using a lung biomechanical model and the estimated elasticity and compared the deformations with the ground-truth 4D deformations using three image similarity metrics: mutual Information (MI), structured similarity index (SSIM), and normalized cross correlation (NCC).

RESULTS

The results show that a cGAN-based machine learning approach was effective in computing the lung tissue elasticity given the end-expiration CT datasets. For the training data set, we obtained a learning accuracy of 0.44 ± 0.2 KPa. For the validation dataset, consisting of 13 4D datasets, we were able to obtain an accuracy of 0.87 ± 0.4 KPa. These results show that the cGAN-generated elasticity correlates well with that of the underlying ground-truth elasticity. We then integrated the estimated elasticity with the biomechanical model and applied the same boundary conditions in order to generate the end inhalation CT. The cGAN-generated images were very similar to that of the original end inhalation CT. The average value of the MI is 1.77 indicating the high local symmetricity between the ground truth and the cGAN elasticity-generated end inhalation CT data. The average value of the structural similarity for the 13 patients was observed to be 0.89 indicating the high structural integrity of the cGAN elasticity-generated end inhalation CT. Finally, the average NCC value of 0.97 indicates that potential variations in the contrast and brightness of the cGAN elasticity-generated end inhalation CT and the ground-truth end inhalation CT.

CONCLUSION

The cGAN-generated lung tissue elasticity given an end-expiration CT image can be computed in near real time. Using the lung tissue elasticity along with a biomechanical model, 4D lung deformations can be generated from a given end-expiration CT image within clinically acceptable numerical accuracy.

Self-channel-and-spatial-attention neural network for automated multi-organ segmentation on head and neck CT images

Accurate segmentation of organs at risk (OARs) is necessary for adaptive head and neck (H&N) cancer treatment planning, but manual delineation is tedious, slow, and inconsistent. A self-channel-and-spatial-attention neural network (SCSA-Net) is developed for H&N OAR segmentation on CT images. To simultaneously ease the training and improve the segmentation performance, the proposed SCSA-Net utilizes the self-attention ability of the network. Spatial and channel-wise attention learning mechanisms are both employed to adaptively force the network to emphasize the meaningful features and weaken the irrelevant features simultaneously. The proposed network was first evaluated on a public dataset, which includes 48 patients, then on a separate serial CT dataset, which contains ten patients who received weekly diagnostic fan-beam CT scans. On the second dataset, the accuracy of using SCSA-Net to track the parotid and submandibular gland volume changes during radiotherapy treatment was quantified. The Dice similarity coefficient (DSC), positive predictive value (PPV), sensitivity (SEN), average surface distance (ASD), and 95% maximum surface distance (95SD) were calculated on the brainstem, optic chiasm, optic nerves, mandible, parotid glands, and submandibular glands to evaluate the proposed SCSA-Net. The proposed SCSA-Net consistently outperforms the state-of-the-art methods on the public dataset. Specifically, compared with Res-Net and SE-Net, which is constructed from squeeze-and-excitation block equipped residual blocks, the DSC of the optic nerves and submandibular glands is improved by 0.06, 0.03 and 0.05, 0.04 by the SCSA-Net. Moreover, the proposed method achieves statistically significant improvements in terms of DSC on all and eight of nine OARs over Res-Net and SE-Net, respectively. The trained network was able to achieve good segmentation results on the serial dataset, but the results were further improved after fine-tuning of the model using the simulation CT images. For the parotids and submandibular glands, the volume changes of individual patients are highly consistent between the automated and manual segmentation (Pearson's correlation 0.97-0.99). The proposed SCSA-Net is computationally efficient to perform segmentation (sim 2 s/CT).

Deformable image registration for radiation therapy: principle, methods, applications and evaluation

Background: Deformable image registration (DIR) is increasingly used in the field of radiation therapy (RT) to account for anatomical deformations. The aims of this paper are to describe the main applications of DIR in RT and discuss current DIR evaluation methods. Methods: Articles on DIR published from January 2000 to October 2018 were extracted from PubMed and Science Direct. Our search was restricted to articles that report data obtained from humans, were written in English, and address DIR methods for RT. A total of 207 articles were selected from among 2506 identified in the search process. Results: At planning, DIR is used for organ delineation using atlas-based segmentation, deformation-based planning target volume definition, functional planning and magnetic resonance imaging-based dose calculation. In image-guided RT, DIR is used for contour propagation and dose calculation on per-treatment imaging. DIR is also used to determine the accumulated dose from fraction to fraction in external beam RT and brachytherapy, both for dose reporting and adaptive RT. In the case of re-irradiation, DIR can be used to estimate the cumulated dose of the two irradiations. Finally, DIR can be used to predict toxicity in voxel-wise population analysis. However, the evaluation of DIR remains an open issue, especially when dealing with complex cases such as the disappearance of matter. To quantify DIR uncertainties, most evaluation methods are limited to geometry-based metrics. Software companies have now integrated DIR tools into treatment planning systems for clinical use, such as contour propagation and fraction dose accumulation. Conclusions: DIR is increasingly important in RT applications, from planning to toxicity prediction. DIR is routinely used to reduce the workload of contour propagation. However, its use for complex dosimetric applications must be carefully evaluated by combining quantitative and qualitative analyses.

Construction of a biomechanical head and neck motion model as a guide to evaluation of deformable image registration

The use of deformable image registration methods in the context of adaptive radiotherapy leads to uncertainties in the simulation of the administered dose distributions during the treatment course. Evaluation of these methods is a prerequisite to decide if a plan adaptation will improve the individual treatment. Current approaches using manual references limit the validity of evaluation, especially for low-contrast regions. In particular, for the head and neck region, the highly flexible anatomy and low soft tissue contrast in control images pose a challenge to image registration and its evaluation. Biomechanical models promise to overcome this issue by providing anthropomorphic motion modelling of the patient. We introduce a novel biomechanical motion model for the generation and sampling of different postures of the head and neck anatomy. Motion propagation behaviour of the individual bones is defined by an underlying kinematic model. This model interconnects the bones by joints and thus is capable of providing a wide range of motion. Triggered by the motion of the individual bones, soft tissue deformation is described by an extended heterogeneous tissue model based on the chainmail approach. This extension, for the first time, allows the propagation of decaying rotations within soft tissue without the necessity for explicit tissue segmentation. Overall motion simulation and sampling of deformed CT scans including a basic noise model is achieved within 30 s. The proposed biomechanical motion model for the head and neck site generates displacement vector fields on a voxel basis, approximating arbitrary anthropomorphic postures of the patient. It was developed with the intention of providing input data for the evaluation of deformable image registration.

Analytical modeling and feasibility study of a multi-GPU cloud-based server (MGCS) framework for non-voxel-based dose calculations

PURPOSE

In this paper, a multi-GPU cloud-based server (MGCS) framework is presented for dose calculations, exploring the feasibility of remote computing power for parallelization and acceleration of computationally and time intensive radiotherapy tasks in moving toward online adaptive therapies.

METHODS

An analytical model was developed to estimate theoretical MGCS performance acceleration and intelligently determine workload distribution. Numerical studies were performed with a computing setup of 14 GPUs distributed over 4 servers interconnected by a 1 Gigabits per second (Gbps) network. Inter-process communication methods were optimized to facilitate resource distribution and minimize data transfers over the server interconnect.

RESULTS

The analytically predicted computation time predicted matched experimentally observations within 1-5 %. MGCS performance approached a theoretical limit of acceleration proportional to the number of GPUs utilized when computational tasks far outweighed memory operations. The MGCS implementation reproduced ground-truth dose computations with negligible differences, by distributing the work among several processes and implemented optimization strategies.

CONCLUSIONS

The results showed that a cloud-based computation engine was a feasible solution for enabling clinics to make use of fast dose calculations for advanced treatment planning and adaptive radiotherapy. The cloud-based system was able to exceed the performance of a local machine even for optimized calculations, and provided significant acceleration for computationally intensive tasks. Such a framework can provide access to advanced technology and computational methods to many clinics, providing an avenue for standardization across institutions without the requirements of purchasing, maintaining, and continually updating hardware.