Eric Bear, Jessica Singh, and Michael Wahrman
The purpose of this paper is to evaluate the use of artificial intelligence in the medical use of particle accelerators for both treatment (radiation therapy) and imaging. We will first discuss the implications of artificial intelligence (AI) on particle accelerators, starting with basic control mechanisms and then moving to complex systems such as neural networking. AI has the potential to increase efficiency and precision in all types of particle accelerators, especially for medical applications. Additionally, it can help lower reliance on trained physicists and also lower the operating costs of accelerators.
Next, we will delve into the field of radiation therapy planning as it pertains to knowledge based systems, artificial intelligence, and treatment planning algorithms. Whether it be in decision making, assessing the validity of predetermined plans, or aiding in the complex process of simulations with dynamic parameters, AI is growing in relevance in the initial stages of treating tumors. In fact, the medical applications of AI in accelerators have increased exponentially over the past decade.
AI can be crucial to obtaining accurate images of the body and properly treating patients. To further understand the role of medical accelerators in proton beam therapy and imaging, we will explore current and future AI technologies. Proton therapy itself offers many advantages over conventional treatments, such as more accurate, effective treatments in addition to fewer side effects.There are concerns that come with artificial intelligence and the use of proton therapy as a whole, but the benefits outweigh the risks. The development of proton therapy in both theory and clinical use will be critical to treating patients in the future.
In the area of particle accelerators in medicine, ion beam radiation therapy requires careful control of beam parameters and targeting. Furthermore, researchers and doctors must be able to image irregular objects and develop detailed therapy plans before treatment. This paper will explore and assess the advancements made in artificial intelligence (AI) technologies to address similar problems. More specifically, the paper will evaluate the past, current, and future developments in AI in radiation therapy as it pertains to planning treatment, targeting irregular objects in various kinds of media, and control of parameters in particle accelerators and ion beams.
The article will first discuss the growing usage of AI in controlling parameters for ion beams and particle accelerators. This section will cover specifics of controlling particle accelerator beams such as basic mechanisms including feedback and feedforward control. Then, past attempts and current projects regarding the use of AI in particle accelerators will be covered. This section will focus mostly on neural networking, a promising area in the field of AI. However, other ideas such as distributed AI and hierarchical AI will be covered. Additionally,
general advancements in AI will be discussed.
Next, the article will discuss the use of AI in Radiation Therapy Planning (RTP). In the field of clinical medicine and radiation therapy, AI is being implemented in decision making processes and continues to evolve with the usage of intricate simulations, the integration of knowledge-based systems, and the accommodation for dynamic, patient-specific parameters. Although AI developments in RTP transform the clinical applications of ion beam radiation therapy, they do not replace the doctors, researchers, and experienced human perspective when assessing irregular objects. Nevertheless, the advancements in RTP that integrate AI provide valuable assets when constantly handling changing parameters and tailoring specific treatment plans to patients. This article will discuss the theoretical purposes of RTP artificial intelligence and include an explanation of various projects, both complete and in progress, using these techniques.
Finally, this article will explore the imaging and targeting of irregular objects in various cases. When developing images for the purpose of medical accelerators, it is vital to take into account factors such as the volume and bounds of the affected region as well as the location in context of the body. In order to tackle such demands, physicians currently use 3D reconstructions from CT imaging and a variety of data visualization and processing algorithms. Through this article, we will explore the effectiveness of these techniques and the future of imaging—especially in terms of artificial intelligence.
In this article, the usage of particle accelerators is integral in radiation therapy. When we mention particle accelerators, we refer to apparatuses that speed subatomic particles to high velocities using either electromagnetic or electric fields. The output of particle accelerators is a focused ion beam with an accuracy and precision based upon specific parameters when steering particles with magnetic fields .
There are several types of particle accelerators, most prominently cyclotrons, betatrons, linear accelerators (linacs), microtrons, and synchrotrons. A cyclotron accelerator is an apparatus that accelerates charged atomic and subatomic particles in an outward spiral or circular path of a magnetic field using an alternating electric field. Meanwhile, a betatron specifically accelerates electrons in a circular or outward spiral path using magnetic induction, or the process of magnetizing the electrons using an external magnetic field. Furthermore, the concept of the cyclotron extends to the microton, which is an accelerator using D-shaped electrodes to create high voltage through a linear structure. Lastly, synchrotrons are specific types of cyclotrons with gradually increasing magnetic field energy strength that keeps the orbit of particles at a constant radius. Linear Accelerators (Linacs), on the other hand, consist of a series of straight line tubing without closed orbits to accelerate particles and they use periodically reversing electric charges to produce a series of peaks and troughs of voltage . All particle accelerators create high-powered beams of charged atomic and subatomic particles known as ion beams. When the article discusses ion beams, it refers to a collection of charged particles. More research and progression in particle accelerator development has led to the use
of proton therapy because it offers stronger penetration to targeted irregular objects. Furthermore, researchers and scientists can control the frequency at which electric fields strip atoms of their electrons or protons when forming a predictable bunch. These parameters play important roles in the foci, accuracy and precision of ion beams, especially as it pertains to radiation therapy.
The diagram shows the schematics of a synchrotron .
The focus of this article is the intersection of computer systems with particle accelerators, ion beams, irregular object imaging, and radiation therapy. When we discuss artificial intelligence (AI), we mean the capability of a computer system to exhibit humanlike intelligence. Moreover, AI deals with the flexibility and adaptability of a computer while perceiving the environment as input to maximize success when achieving a goal. In the field of radiation therapy, artificial intelligence can either be applied to the field of diagnostics and planning treatment or actual therapy with instrumental control or preparation.
Particle accelerators are expensive and demand high maintenance because they are only effective when built according to strict specifications. Whether an accelerator is circular, linear, or combined, accelerators and their magnetic fields along the chamber must be modified according to the attributes of the particles being accelerated. In addition to the strong magnetic forces focusing the beams, particle acceleration must take place in an ultrahigh vacuum. This vacuuming ensures that the particles neither collide with gas nor form unpredictably spaced bunches. The more intricacies used when accelerating particles only raises the extensive use of manpower, time and money. However, these parameters and required technologies can be better analyzed using article intelligence.
More specifically, the instrumentation that characterizes and controls particle accelerator machinery performance and durability must be better controlled. This control stems from the use of a single pulse and a single bunch of uniform particles. Meaning, we must explore the dynamics of the particles in use, their responses to changes in temperature, and how they respond to the target’s rear surface. The immediate application of this technology is applying accelerated particles to an irregular object with efficacy while minimizing damage to healthy tissue surrounding the object. The integration of AI in this component of radiation therapy remains ambiguous, but it is crucial to improving researchers’ and doctors’ control of the dynamics of accelerated ion beams.
Medical Application of Ion Beams and Particle Accelerators
Ion beams are versatile tools with various medical applications, most prominently sputtering in radiation therapy and studying cellular functions in radiobiology. This article will focus on the sputtering, also known as etching, of ion beams onto irregular objects. Sputtering is similar to sandblasting as it uses individual atoms to continually ablate a target. The clinical usage of ion beam sputtering is called ion beam deposition. This process requisites an ion source, ion optics, and a deposition apparatus as well as the occasional usage of irregular object computer analyzers. Furthermore, methodologies of ion beam deposition can be improved with laserdriven systems. Laser acceleration of ions provides more control of bunch and pulse attributes. From a particle accelerator to the use of ion beam deposition and the integration of new technologies, such as lasers, the tangible demands of these technologies remains expensive and resource intensive. Artificial intelligence can lessen the usage of resources with thorough planning, simulations, and knowledge based algorithmic computations .
In medicine, particle accelerators have two applications: clinical diagnostics and radiation therapy. The basic premise of radiation therapy is to effectively direct energy (focused ion beams) to an irregular object target while sparing the surrounding healthy tissue to the best of our ability. This goal is known as local control and drives developments and explorations within the field of particle accelerator usage in medicine. Using more effective and locally controlled particle accelerator ion beams, doctors can treat patients with more accuracy and minimize risk of damaging viable tissue and organs around tumors. On the other hand, particle accelerators are also used in clinical diagnostics, meaning the detection and recognition of irregular objects. This imaging technique is different in method to radiation therapy, but provides equally important resources for clinical uses. Particle accelerators produce radioisotopes emitting x-rays, gamma rays, or positrons as probes. These radioisotopes are used in tandem with instrumentation located outside the patient to image the distribution and interaction of radiation with the biological structures, fluid motion, and constriction of the human body. Both radiation therapy and clinical diagnostics are pertinent to the exponentially growing body of cancer patients. Thus, the exploration of the intersection of AI and particle accelerators in medicine is the frontier that can lead to better patient care.
Proton Therapy and its Benefits
Within the next two decades, the number of new cancer cases will rise to 22 million. In 2016 alone, 595,690 people are anticipated to die from cancer. 39.6% of people will be diagnosed with cancer in their lifetime . As expected life spans increase, the reach of cancer has increased exponentially. Significant steps must be taken in order to combat the disease.
Currently, cancer is primarily treated through immunotherapy, chemotherapy, radiation, and surgery. Ultimately, experts predict that immunotherapy and gene therapy will provide long term solutions and tentative cures for the disease. Though such treatments are promising, progress in those fields takes longer expanses of time — time that most patients cannot afford. Therefore, until systemic treatments can become viable, radiotherapy and surgical removals of tumors will remain the most effective course of treatment.
Approximately half of all cancer patients receive some form of radiotherapy . However, radiation was not always the precise procedure it is today. Rooted in Roentgen’s discovery of x-rays, radiation grew from a hazardous experiment once Coolidge invented the sealed off tube. After dramatic advances in engineering, linear particle accelerators, linacs, have become the primary convention for x-ray accelerators. However, there are significant drawbacks with x-ray accelerator therapy. While in a biological sense x-rays are effective, their absorption is limited to the skin. In fact, x-ray accelerators are unable to penetrate deep within the body without causing serious damage to neighboring tissues. This makes irradiating deep seated tumors especially difficult.
Ernest Lawrence, inventor of the cyclotron, and John Lawrence, physician, were leading pioneers in the field of medical accelerators. They began therapy using neutrons, which inspired the creation of the Fermilab accelerator in 1974. Neutrons have been seen as a favorable form of radiotherapy with their high biological effectiveness and ability to penetrate deep within the body without extensive damage to healthy tissues. However, the dose distribution is extremely poor in regards to neutron therapy. The neutrons disseminate within the tissues after being deposited. This decreases the effectiveness of the treatment, and damages neighboring tissues as well.
This brings us to proton therapy. As of now, proton therapy appears to be the most promising advance in medical accelerators. While there is a diminished biological effectiveness, the targeting is radically better due to the Bragg Peak, an explicit peak on the Bragg curve which occurs immediately before the particle comes to rest. This means that the protons are able to travel deeper within the body, before releasing the radiation. This means that protons actually emit more energy beneath the surface of the body. Therefore with the help of absorbers, irradiating deepseated tumors becomes much easier with proton therapy. Fermi National Accelerator Laboratory developed the first proton therapy accelerator and shipped it off to Loma Linda University Medical Center. It has treated more than 7,000 patients to date.
With each passing day we learn more about particle accelerators and their potential applications in a medical environment. The evolution of medical accelerators has opened up countless opportunities and will hopefully open countless more. Treatments such as proton therapy lead the way in cancer treatment compared to conventional methods. The increased targeting precision, ability to penetrate deep seated tumors, and minimally invasive procedure make proton therapy a strong contender in cancer treatment.
Particle Accelerator Control Issues
The process of controlling a particle accelerator and precisely tuning its beam presents many challenges. For a given accelerator, there can be thousands of separate magnets to control, or inputs, making the beam’s behavior very difficult to predict. Once the accelerator is running, the beam’s path can be influenced by additional nonlinear factors such as sound and shifts in the earth that slightly change the position of magnets throughout the machine . The highest precision possible is needed for accelerators (especially those in the medical field), so many minute factors such as these must be taken into account.
For those reasons, optimizing the behavior of an accelerator becomes a very complex process, especially for humans. Currently, most accelerators around the world require a trained physicist to run properly. These people are very good at their jobs, but are still limited in comparison to computers. The vast amounts of data involved with controlling a beam leads to long time periods spent tuning in order to achieve the desired precision. Often it can take weeks for an accelerator to begin operation after having been shut down. All of this time is lost while scientists wait to use the limited numbers of accelerators across the world.
Additionally, humans are limited in the timescale of their operating ability . People cannot possibly react to events within milliseconds, as a computer can. Also, it can be difficult for people to see patterns or analyze data over longer periods of time, such as hours or days. Having trained physicists to run the machines is another cost to an already expensive project.
Artificial Intelligence and machine learning offer a solution to these problems. By using AI techniques, accelerators will be able to run with higher precision and take much less time to prepare. Systems based on AI, such as neural networking, have the ability to learn from data that is recorded as the accelerator runs. Even though scientists are good at predicting the theoretical behavior of a beam, there are simply too many unpredictable factors involved. As mentioned before, every accelerator is different, and AI systems can adjust to changing environments much faster than humans. When an ion beam goes out of focus, sensors in the accelerator send data to the AI system, allowing it to quickly make the proper adjustments to the accelerator’s magnetic field. An efficient AI system can also decide what information is negligible or unimportant. Humans are relatively good at doing this, and AI systems will need proficiency in this skill since there is so much data coming from sensors in a particle accelerator at any given time . For the medical application of proton therapy, saving time (and therefore also money) is a clear benefit for both patients and doctors.
AI Techniques for Controlling Particle Accelerators Basic Control of an Ion Beam
The control of a particle accelerator’s beam comes down to controlling magnets and electric fields. At its most basic level, tuning an ion beam involves sensing the beam’s location in three dimensions and its momentum in those dimensions. Both the direction of the beam and its level of focus are important. Dipole magnets are typically used to bend the beam’s direction, and quadrupole magnets are used to focus it. After a sensor receives information about the beam (usually in the form of reflected light), the data must be sent to a computer, which then proceeds to analyze . At this point, either a human or the computer makes the necessary adjustments to the system. Most control elements can be sorted into either feedback or feedforward categories.
In feedback control, a sensor records the data at some point on the line. To make an adjustment, magnets before the sensor are tuned, thus changing the properties of the beam where it first was recorded by the sensor. Feedback is generally used to account for random fluctuations in the beam from factors such as noise and mechanical drift . Every system will have unpredictable flaws such as these.
However, systems also contain predictable variables, such the adjustments needed to increase the energy of a beam. In this case, the proper adjustments are already predicted by a computer or human, and feedforward control is more suitable. The values of each variable (such as magnet strength) can be set ahead of time. The disadvantage of a feed forward system is that it will not measure or react to the quality of the beam .
For this reason, a combination of feedforward and feedback loops is the optimal situation
- Feedforward techniques can be used for most control of the beam, and feedback can correct for remaining errors. Thousands of magnets in an accelerator must be precisely tuned to achieve a beam with the desired properties. This can be a daunting task, but AI may be able to make it more manageable.
This diagram demonstrates the two basic ways to control a system .
Here we will delve into how AI is used to optimize the function of particle accelerators in general, while also discussing these implications for proton therapy. AI and machine learning can come in many forms, but machines that run based on these ideas can learn from past experiences and data. This can be very useful for particle accelerators, where huge amounts of information are available for a system to use in optimization. AI systems must analyze the data given to them and react by adjusting magnets in the machine to achieve a desired state of the beam. One promising field in AI is neural networking.
Neural Networks (NN’s) are complex function approximators . They are made up of many connected functions that are assigned different weights in determining an appropriate output. Each of these functions can be adjusted by “training” the system through the use of experimental data. As the machine (in this case an accelerator) is used, the accumulated data is used to fine tune this network of functions, with the ultimate goal being a great level of precision and accuracy. In addition, NN’s can be trained by using data gathered from virtual models and simulations before an accelerator is actually running.
Neural Networking has the potential to be effective in many situations. For example, a physicist could use this system to model a particle accelerator’s behavior, since NN’s are capable of using multiple inputs at once. Also, NN’s great computational speed makes them a candidate for temporarily replacing physics-based models . Since models that apply the basic laws of physics can take a long time to run, NN’s can take their place to find solutions that are very close to the actual answers. And, of course, NN’s could be very useful in operating particle accelerators themselves. This includes both mimicking the routine actions of a humans who work on the accelerators and innovating new control mechanisms .
Diagram showing a conceptual view of how a neural network operates .
So far, neural networking has achieved some success in the field of particle accelerators. The first attempts at using AI in particle accelerators were made in the 1990s. One later instance was a project done by Vista Control Systems and the University of New Mexico. In this project, the team used what they described as a “distributed hierarchical architecture” .
This system was made up of many “controllers” arranged in a hierarchical structure. Each controller had to make decisions about when and what actions would be performed. The system also included “solvers” which could apply basic algorithms when commanded by controllers. Finally, the system incorporated a “Physical Access Layer,” which helped to facilitate communication between actual control events and the software itself.
Other Types of Artificial Intelligence
Of course, other methods of artificial intelligence exist besides neural networking. From 1990 to 2000, many physicists tried to implement computer control in particle accelerators across the world, including CERN in Europe . In their SETUP program, small subsystems of the accelerator were controlled by AI, rather than the entire accelerator. Also, the control only operated before the accelerator was running, during its “set up.” Although the AI program did not control the accelerator in real time, it was still a successful step for future projects.
Another example of AI is Distributed Artificial Intelligence (DAI). This field of study has been investigated for many applications, one of which is particle accelerators. DAI was applied to CERN through the ARCHON (Architecture for Cooperating Heterogeneous Online Systems) project during the 1990s. ARCHON was made up of multiple “intelligent” agents, all of which consisted of many layers. These layers could communicate with each other, allowing the system to designate tasks such as organization and processing. The important part was that each layer could perform its own task but also communicate effectively, allowing for better cooperation .
In reality, most systems that incorporate artificial intelligence use ideas from many different areas, so it is difficult to define a certain system as pure “neural networking” or distributed AI. However, systems that apply at least the basics of neural networking seem to be the most promising for efficiently controlling a particle accelerator . Since their conception in the 90’s, neural networks (and artificial intelligence in general) have gained potential because of technological developments. Advancements in both computer hardware and software have made AI control of accelerators more likely for the future. Computer chips have become smaller, more powerful, and more efficient, while theoretical knowledge of physics and computer programming is constantly increasing.
Artificial Intelligence and Knowledge-Based Computer Systems in Radiation Therapy Planning
Theoretical Applications of Artificial Intelligence in Planning
The most prominent clinical use of artificial intelligence in radiation therapy is in the planning phase, known as Radiation Therapy Planning (RTP). RTP continues to implement computers to support our routine decision making when assessing patient specific cases.
However, our systems have continued to evolve because of our growing capabilities of treatment methods and knowledge based systems. Not only does our access to various types of ion beams and treatment sources continue to grow, but also our ability to continuously handle and account for dynamic patient parameters increases with AI. Using three dimensional simulations and algorithmic flexibility, AI can be integrated into RTP for more effective treatment methods and less use of expensive resources .
When using AI in RTP, we try to implement computers with both knowledge based patient data systems and adaptive algorithms to develop treatment plans. AI is not intended to eliminate humans in the process of detecting, assessing and treating irregular objects. Instead, AI is an aid to indicate what the most favorable treatment method can be. The integral concept of this application of AI is knowledge engineering, the decision making support systems with complex cases, to amalgamate various types of preset data to formulate an output. Specifically, RTP knowledge engineering uses previous patient data sets of locations and attributes of tumors and procedures to produce a plan . This extensive process demands high powered rule based computer programming, but has become essential in the identification of proper treatment plans.
Initially, AI algorithms detect the treatment modality, which is dependent upon the determination of the target’s volume, the dosage desired (with associated risk), and the identification of critical organs surrounding the target that can be limiting factors. Afterwards, the physical method of radiation is decided upon. This process involves an assessment of beam geometry and considerations of financial resources and time. Using these initial algorithms, AI can simulate several different cases of treatment modality and radiation type. As a result, computer systems can formulate an extensive array of prototypical plans.
With these initial plans, AI computers must work in tandem with data sets and humans to assess setbacks and errors in specific cases. These adaptations are made based on available components to treat and changing parameters with the ion beams being used. Lastly, the computer system assesses the favorability of each plan with accommodation for several parameters: dose level, cold spots in predictive beam pathways, cold and hot areas overlapping where treatment should occur, beam dimensions, and changes in the patient’s system. At this point, AI can determine if the tumor should be treated using radiation or not, in which case it recommends surgery, drugs, chemotherapy, or alternative treatment .
Currently, implementation of AI in RTP is not complete because the integration of various AI systems remains a complex problem without a solution. However, several advancements in these systems prove the possible efficacy and revolutionary changes that AI can bring to treatment planning.
An Assessment of Radiation Therapy Planning AI Projects
The University of Washington Project has been in development for over two decades. The initial program was a plan generator that used knowledge bases of past clinical information, patient specific geometry, and a combinatory comparison of the two with several plans. The process works by formulating a general treatment approach along with one or more prototypical plans that build upon the general plan. Furthermore, additions such as dose calculations from simulations have been included to refine plans. However, the system relies heavily on high level descriptions of detected irregular objects to analyze dose computation simulations. As a result, there are region specific results and histograms as well as abstract representations of planned radiation dosage maps. This COMMON LISP program continues to improve, and became the basis for the Prism project at the University of Washington .
The Prism project is a three dimensional radiotherapy planning technique system used on commercialized and standardized platforms. As an improvement to the previous projects at the University of Washington, Prism uses uncommon design features to assure the quality and clinical applicability of the produced plans. The role of AI in this process uses a special tool and support structure for multileaf collimator (MLC) systems, which are bunches of high atomic numbered materials used to block particle beams. As a result, Prism has more flexibility and possibility for expansion. The Prism system includes a three-dimensional analysis and simulation system with a Beam’s Eye View and the flexibility to use any possible treatment geometric attributes and various radiotherapy accelerator ion beam source. Lastly, the system uses a rule-based planning algorithm to find the volume of the target. Prism has been operational in clinical applications since July of 1994, and it has been used at several other locations. Nevertheless, Prism is bent upon simultaneous input from several external users under a common networked data and software base. The Prism system is a substantial step in developing AI tools for tumor and target volume identification, complex and dynamic treatment options, and immediate implementation in computer controlled ion beam sources .
Another AI project in RTP, called CARTES, began at the Technical Research Center of Finland in 1985. This prototypical expert system was intended to aid with decision support in treating various cases of lung cancer. During planning, CARTES will continually critique, assess, and decide upon the consistency of a plan’s efficacy on a specific diagnosis. This decision support system outputs whether the treatment plan should be executed immediately or, conversely, the plan should be modified or re-assessed by physicians because of probable simulation results of treatment. This tool has been applied in several facilities, but remains a check in the execution of radiation treatment planning phases without integration with other AI systems .
One of the more traditional and older AI systems used in RTP is the ONCOCIN project started in 1981 at Stanford University. This program was intended to provide advice to outpatient oncology clinics treating patients with chemotherapy protocols. Although this AI system does not directly apply to particle accelerators, its knowledge engineering uses similar expert rule based algorithms to other AI RTP systems. The benefits of well developed systems like ONCOCIN are the advanced user interface and workstation implementation, which correlates to patient flow charts in traditional clinics. Furthermore, the system makes use of dose modification parameters and time-oriented records from past patient treatments as basis for future cases. ONCOCIN was the one of the first RTP supplements of its kind because it continually built a better basis for decision aid with data collection. However, this project remained just that: a supplement .
While ONCOCIN can apply chemotherapy knowledge protocol and past data to typical situations, another AI system, called ONYX, can be used to amalgamate rule-based knowledge engineering, simulate processes of pathophysiological maps in the body, and utilize decision support theoretical algorithms to evaluate various types of cancer treatment. This process, while it is neither exclusively limited to chemotherapy nor radiation, uses three complex steps to perform its role in the planning phase of treatment. First, the system produces a small number of plans using a patient’s current state and attributes of the irregular object targeted. Consequently, ONYX creates a simulation using the structure and physiology of the relevant surrounding tissues and organs within the body near the target. The result is a predictive set of outcomes of each plan accounting for local control, which is the process of limiting damage to healthy tissue while maximizing destruction of cancerous targets. Ultimately, the system arrives on a satisfying plan from rankings for each plan based on the knowledge gathered in the initial steps. Nevertheless, the program was designed to handle ill specified goals and plan operators with ambiguous consequences (for experimental treatment methods). As a result, the ONYX program’s protocol based knowledge (from ONCOCIN) augments decision making while still necessitating the presence of physicians for analysis of the third step of rankings [7, 11].
Another very specific application of AI in planning is usage of artificial intelligence guiding methods for parameter adjustment in inverse planning, started in the mid 2000s. Inverse planning is bent upon the idea of starting with a plan and modifying it with adaptations based upon additive constraints. The procedure of the program used three main loops to evaluate how an input of physician defined constraints would change possible plans for radiation therapy. Modifications can be made to the prescription dose and volume with respect to changes in critical organs. Ultimately, the program aims to find optimal dosage attributes that remain dynamic and adaptable to other parameters. However, this project is significant for another reason: it was directly compared to the manual method. Unexpectedly, the AI program dosage was not significantly different from the manual method’s dosage. As a result, some AI programs can be effective, but they remain comparable to their manual procedural and human counterparts .
The most recent development in artificial intelligence for treatment planning is the Memorial Sloan Kettering (MSK) IBM Watson Oncology project. The basic premise of the MSK Watson expert system is to aid oncologists by analyzing patient medical records against an extensive data set to produce various options for treatment options. Although this system is not singularly applied to radiation therapy with particle accelerators, its algorithms provide evidence based possible treatments that are detailed and tailored to radiation therapy. With a massive reduction in usage of oncologists’ time to parse research and past patient cases, MSK Watson is one of the most powerful tools in oncology that uses vital knowledge and past records of clinical evidence while overcoming one of the crucial bounds for artificial intelligence: the interpretation and application of input parameters written in plain English. This collaborative program is one of the most modern of its kind and application, which enables the MSK Watson system to access medical journals, textbooks, and the ongoing collection of patient data and text when supporting its decisions for possible plans with evidence. Furthermore, the algorithms can consider administration information for modifications of medication, especially as it pertains to warnings and risks of various drugs. MSK Watson is the most recent development and continues to gain knowledge as an amalgamation of many of the past AI programs used in RTP .
Another currently developing project is artificial intelligence for medical image analysis at imagia. This company is working to better detect and diagnose cancer using traditional image analysis techniques and deep machine learning pattern recognition. The imagia project is working to divide medical image analysis into four steps: detection, segmentation, follow up and classification. More specifically, the program starts by analyzing a set of tumor data with attributes of modality and the type of cancer; this enables the program to detect cancer using an expert system of neural networks and algorithms. Next, the imagia program generalizes the neural networks technique for tumor delineation using propagative learning with simulation and graphical models; the aim of this step of regularization techniques is to increase a network’s ability to accommodate for variation in multimodal cases. Afterwards, the program continues with follow-up examinations of the tumor by means of analyzing different facets of tumor development with algorithms to interpret multiple scans of tumors. Lastly, the system is able to classify stratifying tumors based on its analysis in segmentation and follow-up. As a result, radiologists can diagnose at a quicker pace, assess the time-oriented necessity to treat tumors, and form more complex analyses of tumors than traditional imaging .
Future Work with RTP Artificial Intelligence Systems
There are several other projects in development that use similar ideas of expert, rule-based, knowledge–based, and algorithmic decision-based systems to aid RTP with computers and artificial intelligence. In general, these produce possible plans, protocols or decide upon the validity of plans. These technologies produce humanlike advice for radiation therapy with much faster results and more analytical power. The complexity of RTP without computers also requires extensive amounts of time and money as well as paperwork and record keeping. Computers have become integrated into so many types of research that we must continue to explore this specific application of planning. Ultimately, RTP computer algorithms continue to undergo improvements because we strive to refine our ability to analyze various treatment plans. These improvements integrate larger knowledge–based systems using ever–growing, diverse patient data sets. Subsequently, AI capabilities must maintain three integral foci: 1) the dynamic nature of patient and radiation treatment method parameters, 2) the physical attributes and possible effects of varying dosimetric and geometric radiation properties, and 3) the automatic identification of irregular objects and their surrounding characteristics as it pertains to treatment . In time, we hope to see these improvements and additions to the current works in AI RTP aids.
Previous Applications of Imaging Irregular Masses
While preparing the patient for proton therapy treatment, it is crucial that the accelerator has an accurate and precise image of the affected region. Taking into consideration factors such as the shape, size, and volume of the tumor help the accelerator plan a course of treatment. The imaging ultimately plays an important role in determining dose sizes for precise radiation and thus helps minimize damage to the surrounding healthy tissue.
Imaging of irregular objects and its application in a clinical setting can be divided into two parts: the initial treatment plan and the imaging during the treatment. Initially to develop a treatment plan and obtain a 3D visual, physicians use a reconstruction generally from computed tomography (CT). In the next stage of imaging, different forms of radiation therapy ensure that the treatment remains on course.
CT imaging is generally the more dominant form used for 3D reconstructions and to
determine the volume of the irregular object. Previously, the lack of the sagittal dimension imaging, imaging of planes that divide the body into left and right, made 3D visuals highly inaccurate and unideal for application in radiation and proton therapy. However due to single slice helical CT and multidetector CT (MDCT) scanners, 3D reconstruction has made giant leaps of progress.
Helical CT scans are gathered by a rotating x-ray beam. Once the patient is in the gantry the x-ray beam follows a helical path that results in a three dimensional data set. Hence, spherical CT’s are able to create enhanced 3D images. The more recent MDCT scan take these advances one step further. The revolutionary MDCT, unlike previous CT scans, has a two dimensional instead of a linear array of detector elements. This enables the MDCT to gather multiple slices simultaneously and thus greatly increase the acquisition speed of the CT scanner. Furthermore, MDCT scanners gather volume data, rather than slice data. This means that the MDCT develops 3-D images using the two dimensional detectors in contrast to the 2-D images developed by the linear array of detectors in a regular CT scan. This makes sure that the image obtained is isotropic and therefore ideal for 3D view in medical accelerator applications .
Despite the improvements that come along with the new CT imaging systems, there are some difficulties. While the rapid accumulation of data from MDCT scans can be seen as a positive diagnostic tool, it can also be a hinderance. The sheer number of images may be impractical from a radiologist’s perspective and increase costs extraordinarily .
Once the scans are gathered, the physicians then proceed to determine the treatment dosage. The dosage that the accelerator will administer at each point of the treatment is determined by the respective volume of the region. For data visualization volume rendering techniques such as maximum intensity projection (MIP), minimum intensity projection (MinIP), and shaded surface display (SSD)/volume rendering (VR) are applied.
MIP is an algorithm for data visualization that is able to detect hyper dense regions. From its sensitivity to intense structure, MIP creates a bidimensional image. Conversely MinIP is a method that detects low density structures to produce a bidimensional image. SS VRT provides a completely different visual. It makes a 3D representation that can be viewed from any perspective .
FIGURE A: CT Scan / MIP derived Image
FIGURE B: CT Scan/ MinIP Image
FIGURE C: SS VRT Image 
To further assist in the treatment process the centerline is sometimes also extracted to develop a more comprehensive image of the tumor and clearer path for the accelerator to follow. By using a morphology guided level set model a centerline extraction can be performed. An important factor of this technique is that the extraction is the main source of the imaging for the physicians. First they begin by understanding the structural patterns of the tubular object. To determine the centerline in the object, they assume the centerline is the path where there is minimal flux in a gray image (determined using Eikonal equation). 
Application of Imaging During Treatment
Once the initial imaging and the treatment plan has all come together, the treatment itself requires a second layer of imaging. During the radiation therapy there are a variety of methods that make sure the accelerators follow the path that has been predetermined for them. In fact, there is new technology that allows continued adaptability that repeatedly adjusts for the basic breathing movement of patients without restricting them to a stereotactic frame.
3D radiation therapy techniques include intensity modulated radiation therapy (IMRT), stereotactic radiosurgery (SRS), and gamma knife/cyberknife. Image guided radiation therapy (IGRT) helps detail the tumor logistics for the different radiotherapy methodologies. It is able to use the forms of imaging we discussed earlier to create 3D coordinates for the treatments to follow. This allows IGRT to confirm the location and the shape of the tumor prior to and during the treatment.
In particular IMRT is a form of radiotherapy involving computer-controlled linear accelerators that targets particular locations within the tumor. Furthermore through handling the intensity and duration of the radiation beam IMRT is able to exercise precise control. SRS uses gamma rays or x-rays to treat small brain tumors. In order to make sure the accelerator is unable to deviate from the exact positioning stereotactic frames are used hence holding the patient absolutely still. Finally, the cyberknife provides the most adaptable form of control. Without using body frames or fiducial markers, the cyberknife is able to treat the patient with ease and can accommodate for any slight movement caused by breathing and other biological processes. The real time adaption from the computer systems is a result of hundreds of intersecting beams pointed towards the tumor that increase the malleability of the machine. In fact, the delivery of the radiation is determined to be so accurate that it is safe for spinal and other CNS tumor removals.
Overall the imaging process has been making continual improvements, and though there is human involvement in the process, there have been steps towards automating the procedure for the sake of effectiveness and efficiency.
Concerns and Future Work
While the application of artificial intelligence techniques to medical accelerators for proton beam therapy is a novel and revolutionary idea, the appeal does not fully transition from a lab to a clinical atmosphere. As far as the theoretical logistics are considered, though intriguing in a clinical setting, there are many difficulties that arise. Therefore, it is important to analyze both the problems associated with the actual application of the treatment and the integration of it in hospitals.
Medical Side Effects of Proton Beam Therapy (PBT)
Speaking from a strictly medical perspective, proton beam therapy has proven itself to be a viable method of radiation therapy. Radiotherapy comes with significant risks; however, using proton therapy has shown a general trend towards preserving larger concentrations of healthy tissue and an increase in targeting of tumors precisely. This leads proton beam therapy to be a relatively safer and more effective option for cancer patients as far as treatment is concerned.
However, despite the general increase in safety there are still side effects that can be associated with proton beam therapy—even if these side effects are less likely and often times do not have as intense repercussions. Some side effects include: fatigue, eating and digestion problems, headaches, loss of hair around the region of the body that was treated, redness of the afflicted region, and soreness of the part of the body that was treated. However, the concerns and side effects do vary based on the part of the body that is being treated and other factors.
One particular example we can explore to better understand the possible results from proton beam therapy and its limitations is prostate cancer. The exclusivity of proton beam medical accelerators and their scarcity has resulted in no universally approved or used method for in vivo determination of proton range and dose determination. This makes it difficult for individual physicians to develop treatment plans since there are no precedents for them to follow. However, to address this concern there is research being done in PET scanning and other similar techniques to determine proton dose and range.
It is important to realize that overall the medical efficacy of the proton beam therapy is one that is commendable and comparatively better than previous techniques. However, there is a standard list of side effects as with any kind of high intensity treatment. Additionally, it should be noted that the larger theoretical issues that come are a result of the novelty of the treatment. It has only been present for a limited period of time and therefore there are many standards and tests that are needed for a full treatment that have not been determined. Furthermore, there is current work being done in those fields to help develop the surrounding tests and logistics to make the overall proton beam therapy run smoother and more convenient in future applications.
Economic Concerns of Proton Beam Therapy and Medical Accelerators
Linear medical particle accelerators are widely used for different purposes in hospitals. They appear as a common commodity; however, cyclotrons used for proton beam therapy have not been able to reach that same level of integration. As we move away from the theoretical development of proton beam therapy and explore the real world clinical integration of the treatment it is crucial that economic concerns such as expenses are taken into consideration. The extraordinarily high cost of the proton beam therapy accelerators has resulted in them being a limited resource and therefore ineffective as a standardized treatment since they are not easily accessible.
Through an in-depth cost analysis of PBT it becomes evident that it is not financially viable for many hospitals and therefore inconvenient. The construction of a proton beam facility is approximately around 180 million dollars. Additionally, compared to currently used radiotherapy facilities, the maintenance is extraordinarily higher. “The Institute of Clinical and Economic Review (Boston, MA) estimated the lifetime costs and quality adjusted life expectancy for IMRT to be $45,591 and 13.81 years, and $72,789 and 13.7 years for PBT” . Therefore clearly showing that the current method of imaging modulated radiation therapy is significantly more cost effective and since most hospitals already have access to such facilities there is no need for additional building costs as there is with the proton beam therapy facilities.
Artificial Intelligence Concerns
Artificial Intelligence is the way of the future and increasingly we see scientists becoming more and more involved in creating neural nets that can enable deep learning and make our current machines more effective. In fact, in the case of particle accelerators we could see increased accuracy in imaging and decreased costs for maintenance if there is increased integration of article intelligence in features such as beam line tuning.
Basic machine learning is a form of artificial intelligence where the systems follow a series of well defined steps to arrive at a conclusion. Deep learning is more ambiguous and involves the system drawing from a variety of knowledge that it has compiled over time. The overall process seems rather fascinating and a viable option of making proton beam therapy and medical accelerators in general more effective. However, there are large concerns as to the role of the surrounding scientists and their inability to understand the decisions the machine comes to.
There are several concerns that arise as deep learning and complex neural networks become the future for artificial intelligence. If there are malfunctions, then it will be difficult for the scientists and physicians to decipher how the machine came to its conclusion and therefore they will be unable to rectify whatever the flaw was.
Though there are many concerns that are both valid and important to take into consideration, we should not let these concerns hinder progress in the field of proton beam therapy and the artificial intelligence integration into medical accelerators. The majority of the theoretical concerns are because the PBT is fairly new and therefore not well developed. With time and work, progress will be able to address many of the concerns and make PBT a viable option for hospitals everywhere.
In the last few decades, artificial intelligence-based programs have become more and more integrated into our technology, especially in the fields of particle accelerators and their application in medical therapy. Many attempts have been made to control and tune particle accelerators using AI, with the ultimate goal being an increase in machine efficiency and precision. In addition, the reliance on trained physicists to run such complex machines is likely to go down. With AI’s ability to learn from data and simulations, it promises to be an incredibly effective tool some day in the future. Though many forms of AI have been used, neural networking seems to be the most encouraging. Because of its ability to take in numerous inputs and flexibility in computing, it could be the most ideal system for particle accelerators.
Radiation Therapy Planning is also evolving with its integration of artificial intelligence because computer systems are improving with versatility and complexity. AI planning aids integrate three dimensional simulations, flexible algorithmic analysis, and assess patient specific cases with previous data sets and dynamic parameters. Although these aids do not remove the human from the planning process, knowledge and rule-based engineering bolster clinical capabilities to provide various options. Furthermore, these projects are in development, some of which are at the University of Washington, IBM, Imagia, Stanford and the Technical Research Institute of Finland. Nevertheless, these programs have yet to converge in a manner conducive to an extensive planning system to not only assess treatment options, produce procedures, assess the validity of plans, and adapt to the changes of patient cases. These combination systems have yet to become a reality, but the programs developed so far have greatly aided in high-powered computational analysis.
In the future, it is anticipated there will be development in both AI medical accelerators and radiation therapy. Working in conjunction, the two fields show great promise for diseases such as cancer. Currently, the cutting-edge proton beam treatment uses a cyclotron and CT imaging. However, it is crucial that artificial intelligence methods for identifying the shape and volume of an irregular object are developed. This way, human error can be taken into account so that the accelerator targeting the tumor can do so with greater precision and accuracy. Despite the concerns that neural networks will become too complex to fully understand, steps must be taken to gauge the effectiveness of artificial intelligence in medical accelerators.
- “Systems and Fact Finding,” CSUSB. Aug. 2016 http://www.csci.csusb.edu/dick/cs557/a1.html
- A.L. Edelen, S.G. Biedron, B.E. Chase, D. Edstrom, S.V. Milton, and P. Stabile, “Neural Networks for Modeling and Control of Particle Accelerators,” IEEE Transactions on Nuclear Science, April 2016.
- Accelerators for Society Project, “Accelerators for Society,” TIARA Project, Aug. 2016, http://www.accelerators?for?society.org/about?accelerators/index.php?id=21.
- B. Dotson, “How Particle Accelerators Work,” Energy.gov, Jun. 2014, http://energy.gov/articles/how?particle?accelerators?work?.
- Burrill, J., Z. Dabbagh, F. Gollub, and M. Hamady. “Multidetector Computed Tomographic Angiography of the Cardiovascular System.” Postgraduate Medical Journal 83.985 (2007): 698?704. Web.
- Cole, Francis T. “Medical Accelerators.” The British Medical Journal 2.4679 (1950): 616. Web.
- E. H. Shortliffe, “ONCOCIN Annual Report,” Departments of Medicine and Computer Science at Stanford University with support of National Library of Medicine, 1982?1983.
- Fink, J. R., M. Muzi, M. Peck, and K. A. Krohn. “Multimodality Brain Tumor Imaging: MR Imaging, PET, and PET/MR Imaging.” Journal of Nuclear Medicine 56.10 (2015):1554?561. Web.
- H. Yan, F.F. Yin, C. Willett, “Evaluation of an artificial intelligence guided inverse planning system: clinical case study,” Department of Radiation Oncology, Duke University Medical Center, April 2007.
- I.J. Kalet, J.P. Jacky, M.M. Austin ?Seymour, S.M. Hummel, K.J. Sullivan, J.M. Unger, “Prism: a new approach to radiotherapy planning software,” Radiation Oncology Department, University of Washington, Sep. 1996.
- I.J. Kalet, W. Paluszynski, “Knowledge Based Computer Systems for Radiotherapy Planning,” Radiation Oncology Department and the Computer Science Department, University of Washington, 1990.
- IBM, “IBM Watson for Oncology,” IBM, 2016, http://www.ibm.com/watson/watson?oncology.html.
- Imagia Cybernetics Inc., “Artificial Intelligence for Image Analysis,” Imagia Cybernetics Inc., 2016, http://imagia.com/#product.
- Mabray, Marc C., Ramon F. Barajas, and Soonmee Cha. “Modern Brain Tumor Imaging.” Brain Tumor Research and Treatment Brain Tumor Res Treat 3.1 (2015): 8. Web.
- Maher, Michael M., Mannudeep K. Kalra, Dushyant V. Sahani, James J. Perumpillichira, Stephania Rizzo, Sanjay Saini, and Peter R. Mueller. “Techniques, Clinical Applications and Limitations of 3D Reconstruction in CT of the Abdomen.” Korean J Radiol Korean Journal of Radiology 5.1 (2004): 55. Web.
- M. Dehler, “Real time control of beam parameters,” Paul Scherrer Institut, 2007, https://cds.cern.ch/record/1100540/files/p331.pdf?.
- M. Pieck, “Artificial Intelligence Research in Particle Accelerator Control Systems For Beam Line Tuning,” Los Alamos National Laboratory, 2006, http://permalink.lanl.gov/object/tr?what=info:lanl?repo/lareport/LA?UR?08?06221\
- Perandini, Simone, N. Faccioli, A. Zaccarella, Tj Re, and Pozzir Mucelli. “The Diagnostic Contribution of CT Volumetric Rendering Techniques in Routine Practice.” Indian Journal of Radiology and Imaging Indian J Radiol Imaging 20.2 (2010): 92. Web.
- P. Kolari, J. Yliaho, K. Näriäinen, S. Hyödynmaa, A. Ojala, J. Rantanen, N. Saranummi, “CARTES — A Prototype Decision Support System in Oncology,” Medical Engineering Laboratory at the Technical Research Centre of Finland, Tampere, Finland, Department of Radiotherapy at the Tampere University Central Hospital Pikonlinna, Finland, and the Laboratory for Information Processing at the Technical Research Center of Finland, Helsinki, Finland, 1991.
- P. Skarek, “Multi?Agent Cooperation for Particle Accelerator Control,” CERN, 1996, https://core.ac.uk/download/files/410/11857513.pdf
- Picture Source
- http://4.bp.blogspot.com/?EK?Tzu1XyrQ/UTWjCBM8ApI/AAAAAAAAAFw/iM8Khq1htww/ s400/neural?network.png
- Sabado, Maurice M. “Medical Applications of Accelerators.” Supercollider 3 (1991):639?48. Web.
- Santamaría?Pang, A., C. M. Colbert, P. Saggau, and I. A. Kakadiaris. “Automatic Centerline Extraction of Irregular Tubular Structures Using Probability Volumes from Multiphoton Imaging.” Medical Image Computing and Computer-?Assisted Intervention – MICCAI 2007 Lecture Notes in Computer Science: 486?94. Web.
- S. Braccini, “Particle Accelerators and Detectors for medical Diagnostics and Therapy,” Science Faculty of the University at Bern, April 2013, https://arxiv.org/pdf/1601.06820.pdf.
- S. Petrovic, N. Mishra, S. Sundar, “A novel case based reasoning approach to radiotherapy planning,” Automated Scheduling, Optimization and Planning Research Group School of Computer Science and IT, University of Nottingham, and Department of Oncology, Nottingham University Hospital NHS Trust Nottingham, 2011.
- Stock, Richard, and Nelson Stone. “Combined Modality Therapy With Brachytherapy and External Beam Irradiation in the Management of Localized Prostate Cancer.” Advances and Controversies Management of Prostate Cancer (2004): 233?55. Web.
- Wisenbaugh, Eric S., Paul E. Andrews, Robert G. Ferrigni, Steven E. Schild, Sameer R. Keole, William W. Wong, and Sujay A. Vora. “Proton Beam Therapy for Localized Prostate Cancer 101: Basics, Controversies, and Facts.” Reviews in Urology. MedReviews, LLC, 2014. Web. 05 Sept. 2016.
- W. Klein, G. Luger, E. Olsson, and C. Stern, “An Intelligent Control Architecture for Accelerator Beamline Tuning,” Vista Control Systems and University of New Mexico, 1997, ?https://www.aaai.org/Papers/IAAI/1997/IAAI97?203.pdf
- Zhu, Xuping, and Georges El Fakhri. “Proton Therapy Verification with PET Imaging.” Theranostics 3.10 (2013): 731?40. Web.
- Gibbs, Keith. ?Synchrotron?.Digital image. ?Schoolphysics?.Keith Gibbs, 2016. Web. 15 May 2017.
Dr. Sandra Biedron, Particle Accelerator Laboratory and Electrical & Computer Science Department, Colorado State University
Brief Biographies of Authors
Eric Bear – Eric is a junior in high school at Colorado Academy in Denver, Colorado. Eric is passionate about all sciences and mathematics, especially physics and calculus. Over the past two years, he has developed a deep interest in computer science. Over the summers of 2016 and 2017, Eric has worked and will work at the Air Force Research Laboratory to apply his programming knowledge to computer engineering and physics applications. Eric plans to pursue engineering, computer science and mathematics in college and beyond. Outside the classroom, Eric is a Boy Scout with Troop 130 of Golden, an avid soccer player for his school, a mock trial expert witness, and an involved actor in his school’s Conservatory of Theatre program. In his free time, Eric tries to get outside as much as possible, whether it be hiking, camping, climbing, or hiking.
Jessica Singh – Jessica is a senior at Irvington High School in Fremont, CA. Jessica is extremely interested in the core sciences such as mathematics and physics. Additionally, she has cultivated a passion for computer science and computational biology over the last four years through different experiences. In the future, she hopes to major in computational biology. Jessica is additionally an avid reader, enjoys playing tennis for her school team, and is a nationally ranked debater.
Michael Wahrman – Michael is a senior in high school this year at Cranbrook Kingswood in Bloomfield Hills, MI. Currently, physics interests him the most of any school subject, but he enjoys art classes and writing for the school newspaper. He plans to major in physics in college but is still considering medical school for graduate studies later down the road. Outside of the classroom, Michael loves playing basketball and challenging his friends to games of table tennis.