Quantitative Image Analysis
While TNM staging (stage I-IV) has been shown to effectively stratify patients for therapeutic decision making, non–small cell lung cancer (NSCLC) patients with the same TNM stage often have very different clinical outcomes, illustrating the inadequacy of TNM stage as the lone treatment determinant. Hence, there is a clear need for additional patient and tumor information to drive clinical care toward further personalization beyond TNM stage.
We have demonstrated that, in stage III NSCLC patients, the combination of quantitative image features (based on pre-treatment CT and PET scans) with conventional prognostic factors (such as TNM stage and tumor volume) can improve classification of patients into low- and high-risk groups even though these patients all look very similar clinically. Patients stratified by a PET-based model into the low-risk group had a 60-70% probability of surviving 3+ years; those in the high-risk group had a 20% probability of surviving 2 years. Furthermore, although recent RTOG trial results indicate that, on average, radiation dose escalation is not advantageous and may actually be harmful, we have demonstrated that we can use these same imaging-based models to identify patients who will or will not benefit from dose escalation. Thus, this area of research (known as “radiomics”) has tremendous potential to provide outcome models that could give accurate and reliable outcome predictions that would be of immense value to patients, their caregivers, and clinical staff.
Additionally, these models could potentially be used to determine the optimum treatment for individual patients. Radiomics also has potential for improving the utility of future clinical trials, as it helps separate patients into low- and high-risk cohorts prior to the actual treatment. All of this additional information is essentially free, as the analysis is based on images that are already being taken as part of the patients’ routine clinical care. Our current research focuses on the use of quantitative image features to understand individual patient’s response to treatment and the optimization of these features (to maximize information content and reproducibility).
Use of CT to quantify radiation-induced esophagitis
Incidence or severity of esophagitis is the primary or secondary endpoint in almost all clinical trials of treatments for lung and esophageal cancer. Esophagitis can negatively affect treatment efficacy by preventing dose escalation to the tumor in addition to causing an interruption in treatment. The determination of esophagitis grade is highly subjective with many sources of variance. One of the physical manifestations of clinically significant esophagitis is swelling within the esophagus. We have shown that radiation-induced esophageal expansion is a continuous, objective measure of esophagitis by strongly linking expansion to patient symptoms. Furthermore, using this quantification of esophagitis overcomes the variance and subjectivity of traditional scoring methods. We have developed a method using CT imaging with deformable image registration and in-house software to calculate esophageal expansion at the voxel level. This method gives an accurate quantification of the extent and spatial location of esophageal swelling. We have substituted esophageal expansion for esophagitis grade as an endpoint in prediction modeling. This gave an improved model performance allowing more precise esophagitis prediction based on radiation therapy plan, before treatment begins.
Currently, we are using this method to better understand radiation dose-response in the esophagus. This knowledge to can help to prevent severe esophagitis, thereby improving treatment efficacy. In addition, we are investigating esophageal expansion as a measure of difference in biological response between photon and proton irradiation, if such a difference exists.
Virtual endoscope tracking for video-CT image registration
Visual endoscopy is an important imaging modality for many cancer patients, but it cannot be fully integrated with modern radiotherapy until a method is developed to register endoscopic video to CT. We are investigating the use of virtual endoscopic images to achieve this using only image processing techniques. The goal is to create a tracking/registration approach that can be used with current endoscope designs.
At present we have started to evaluate our approach with phantoms. Our method can successfully track the position and orientation of an endoscope in a recorded video, and it provides accurate mapping of spatial locations from video to CT. It needs to be made more robust to unfavorable camera motion and false minima in the coordinate search, but these results show that this method can serve as a foundation for an endoscopy/CT registration framework that is clinically useful and achievable with standard equipment.
A preclinical study investigating the impact of a magnetic field on radiation-induced toxicity
The clinical use of the new Elekta MRI-LINAC introduces some interesting radiation dose questions, including gross dosimetric effects of the magnetic field (which can be modeled using Monte Carlo) and potential radiobiological/physiological effects that are, at present, unknown. Our general objective is to use pre-clincial studies to demonstrate the radiogiological consequences of magnetic-field-induced dose pertubations. We have already demonstrated, using Monte Carlo simulations, the suitability of a Co-60 irradiator / 1.5T magnetic field for these experiments. Our current work is focused on the evaluation of this combination for pre-clinical studies, including the use of micro-CT imaging to assess pneumonitis.
Radiation therapy has been shown to be a cost-effective therapy for curative cancer treatments in low- and middle-income countries (LMICs). For many malignancies, including locoregionally advanced carcinomas of the uterine cervix and head/neck, radiation therapy is the only effective treatment modality; there is an extremely poor prognosis for these patients, with no reasonable curative alternative if radiation therapy is not available. The IAEA estimates that there is a shortfall of 5-10,000 treatment units. Furthermore, it is estimated that by 2020 there will be a deficit in radiation therapy staff in LMICs of more than 50,000 full-time equivalent staff, including almost 10,000 medical physicists.
This project is one of the steps towards creating a fixed-beam LINAC that could be suitable for LMI countries, where 75% of cancer patients do not have access to radiation therapy (data source given in footnote a). A fixed-beam LINAC could, potentially, be manufactured at a fraction of the cost of current LINACs. In addition to this global view, the following patient groups could have specific clinical benefits from vertical (seated) treatments. This is important, as these groups are likely the ones for whom we will be able to get initial clinical experience in treating in this position (in our own clinic):
- Thoracic patients. Positioning the patient in a vertical (seated) position increases lung volume and reduces lung motion, compared with lying down (demonstrated as part of a previous Varian MRA project). If treated in this position, there would be a corresponding reduction in the amount healthy lung that is irradiated, potentially reducing treatment side effects - thus having a similar benefit to deep-inspiration breath-hold techniques, but more easily tolerated by patients.
- For many patients, being seated in a vertical (or slightly inclined) is more easily tolerated, especially for those with compromised lung function or excessive saliva accumulation. For head and neck patients, who are immobilized in a mask which is locked to the treatment couch, this can be a safety concern.
- Young Hodgkin’s lymphoma patients, where it is important to keep the breasts and heart out of the radiation field to minimize the risk of breast cancer induction and coronary artery disease. The breasts and heart move down and away from the treatment field if the patient is in a more vertical position. Thus, a more vertical treatment approach could minimize the risk of secondary breast cancer and heart disease later in life.
Thus, this project may benefit a significant (but unknown) number of patients in the USA, and a much larger group of patients in LMI countries. This could make a significant impact on world health. Finding appropriate technologies to treat these patients is a current priority of several UN agencies (e.g. IAEA)
A quantitative approach to comparing treatment units for use in a low-resource setting.
Cobalt-60 units and linear accelerators are available to meet the teletherapy need around the world. There are many discussion papers comparing the types of treatment unit for use in a low-resource setting. In this project we attempted to quantify the impact of machine choice, power infrastructure, machine downtime, and treatment technique on relative daily patient throughput. This quantitative analysis elucidates more clearly the advantages and disadvantages of each machine operating under various levels of resource availability.
Data from patient treatments plans, peer reviewed studies, and international organizations were combined to compare the relative patient throughput of linacs delivering four treatment techniques (conformal, step-and-shoot Intensity-Modulated Radiation Therapy , dynamic IMRT, and Volumetric-Modulated Arc Therapy) and cobalt teletherapy units (conformal only) under varying power conditions and maintenance support. Data concerning power outage frequency and duration and downtime characteristics of the teletherapy units were used to model teletherapy operation in low-resource settings. The quantitative analysis showed that a decrease in average daily throughput was seen for linacs due to lack of power infrastructure and for cobalt units due to the limited and decaying source strength. Under conformal treatment techniques, delivered with multi-leaf-collimators, we found an equal average daily patient throughput when an average of 2.34 hours of power outage is experienced per 10-hour working day. Relative throughput of linacs delivering step-and-shoot IMRT, dynamic IMRT, VMAT compared to linac conformal techniques saw decrease in throughput of 32, 24, and 20%, respectively. Additionally, the effect of block changes, needed when MLCs are not available, was studied for cobalt conformal treatments. Throughput was decreased 14 and 37% when 1 and block changes were needed per patient for fraction, respectively.
In summary, in scenarios of many power outages, cobalt-60 teletherapy units are best suited for implementation. However, each scenario and region for implementation is unique, and many additional factors must be considered (such as service response time, availability of parts, number and training of staff, interest of the individual clinic, projected patient population, and many others). This is a very complicated question and overall conclusions are difficult to draw.
Image segmentation and deformable image registration
Multi-atlas segmentation for radiation treatment planning
Modern radiation therapy techniques require clinical specialists accurately define the targets and the concerned organs-at-risks for radiation treatment planning. Traditional manual contouring is labor-intensive and time-consuming. In addition, manual contouring introduces intra- and inter-observer variations that are caused by different clinical experience, training of the specialists, and quality of available medical images. As a result, computer-aided robust automatic segmentation becomes increasingly important to reduce contouring time and create more accurate and objective standardized contours.
Over the past several years, we have been dedicated to the development of multi-atlas segmentation to facilitate radiation treatment planning. The multi-atlas segmentation used deformable image registration to deform multiple atlas contours to a new image. We developed a tissue appearance model built into the simultaneous truth and performance level estimation (STAPLE) algorithm for efficient contour fusion generating auto-segmented contours for the new image. On the basis of the multi-atlas segmentation development, we have created a software tool, Multi-Atlas Contouring Service (MACS), which interfaces with the Pinnacle treatment planning system, for clinical use of auto-contouring. We have developed and validated the following atlas structures used in MACS: parotid glands, mandible, submandibular glands, brachial plexus, esophagus, and cardiac structures (heart, chambers, and great vessels). Currently we are expanding the MACS atlas database and continue to improve the multi-atlas segmentation algorithm.
Optimal atlas selection for atlas-based segmentation
Atlas-based segmentation (ABS) has become an important tool to automatically delineate anatomical structures in modern radiation therapy. Deformable image registration plays a key role in ABS to transform the atlas structures onto the new image. Nevertheless, the high variations in intensity, contrast, and geometry across different patients pose a great challenge for deformable image registration. A bad atlas usually introduces large registration discrepancies leading to inaccurate segmentation. With the progress of radiation treatment, images with manual contours become more frequently available in the database with the potential to serve as atlases to delineate structures for new patient images. We developed a local atlas selection strategy to identify from the database on-the-fly a best atlas or a set of optimal atlas candidates that are the most similar to the new image to perform atlas-based segmentation. In this project, we investigated robust and reliable image features, such as 3D Gabor texture features and geometric moment invariants features for local feature matching to identify optimal atlas candidates.
Uncertainty analysis of deformable image registration for
During a head and neck IMRT treatment course, significant changes in disease and normal anatomy can be observed, such as the regression of primary tumor and nodal disease, normal glands and mucosa alterations, and weight loss. Continuing to deliver the original IMRT plan may cause significant toxicity to normal tissues. Adaptive radiotherapy has emerged as an approach to correct for these daily tumor and normal tissue variations through online or offline modification of original IMRT plans. The core technique to enable a practical adaptive radiotherapy is the deformable image registration. However, currently there is no effective approach to evaluate the uncertainty of deformable image registration and the related dosimetric impact in adaptive radiotherapy. The main reason may be the lack of ground truth of deformation to validate the deformable image registration algorithms.
This on-going project creates a data model by learning the shape and appearance variations from patients receiving radiation treatment before, and generates artificial images from the data model for deformable image registration validation. By using Monte Carlo simulation, we produce a large database for the creation of registration uncertainty models. Finally, by adapting the general registration uncertainty models to a specific patient through classification, we are able to perform patient-specific uncertainty analysis.
Improve CBCT image quality through deformable image
The advent of in-room cone-beam computed tomography (CBCT) provides an accurate representation of patient anatomy prior to treatment delivery, allowing for daily patient setup, evaluation of changes in the tumor and critical structures, treatment re-planning, and dose verification/accumulation in adaptive radiotherapy. However, 3D-CBCT lacks explicit representation of patient respiratory motion and usually has poor image quality and inaccurate CT numbers for target delineation and/or adaptive treatment planning. In-room 4D-CBCT image acquisition is still time consuming and suffers the same issue of poor image quality. To overcome this limitation, we developed a computational framework to digitally synthesize high-quality 4D-CBCT images using the prior knowledge of motion and appearance learned from the planning 4DCT. A patient-specific respiratory motion model constructed from the planning 4DCT the image content of the planning 4DCT was spatially mapped onto the daily CBCT using deformable image registration. The synthesized 4D-CBCT images possess explicit respiratory motion of the patient and their image quality is comparable with that of conventional CT images. In this project, we also investigated a density correction approach to address large deformations in lung and a new regularization scheme for deformable registration to address motion discontinuities along different types of tissues such as lung sliding issue.