MJM MedTalks
A Conversation with Dr. Shirin Enger

S01E02

Nadia Blostein for the McGill Journal of Medicine1
Published online: October 28, 2022

1McGill University
mjm.med@mcgill.ca

Abstract

McGill Journal of Medicine (MJM) MedTalks is a Podcast series where members of the McGill Faculty of Medicine and Health Sciences are interviewed on topics related to career, research, advocacy and more. The aim of MedTalks is to open a space where faculty members can share information and advice for trainees in healthcare and medical sciences. In this episode, MSc candidate and MJM Podcast Team member Nadia Blostein interviews Dr. Shirin A. Enger, medical physicist and Associate Professor at the Gerald Bronfman Department of Oncology & Medical Physics Unit. This interview focuses on the interdisciplinary and translational nature of medical physics research, the workflow of radiotherapy, and concludes with a list of open-source initiatives that Dr Enger’s group has participated in, including the McMed Hacks Series on Machine Learning and Medical Imaging. The show notes include a transcript of the podcast, a more detailed content overview, glossary of important terms and resources and references. This episode was recorded and edited by MJM Podcast Team member Nadia Blostein with input from the entire MJM Podcast Team and transcribed by Susan Wang.

     

Content overview

0:18 Introduction;

2:45 Clinical vs. research medical physics;

9:35 The interdisciplinary, collaborative and translational nature of medical physics research including the workflow of radiotherapy and the differences between outside-in external beam radiation and inside-out brachytherapy;

11:18 Radiation therapy vs. nuclear imaging medical physics;

15:07 The brachytherapy pipeline that Dr. Enger’s group is seeking to automate and merge into one unified software system;

25:38 How machine learning has enabled the acceleration of brachytherapy dosimetry map prediction through the development of the RapidBrachyDL algorithm;

35:28 The open source initiatives that Dr. Enger’s group has been a part of: RapidBrachy MCTPS, an open source treatment planning system that the research group is developing and McMed Hacks, a summer workshop series on machine learning and medical imaging spearheaded by Dr. Enger’s students Yujing Zou and Lucas Weishaupt.

37:07 Brief overview of a creative technique called needle shielding;

39:39 Outro

Glossary

Links and papers

Transcript

0:18 Meryem Talbo (MT): Welcome to a new episode of the McGill Journal of Medicine, the MedTalks series. In this series, we interview members of the McGill Faculty of Medicine and Health Sciences on topics related to career, research, advocacy and more. The aim of our MedTalks is to open a space where faculty members can share information and advice for trainees in healthcare and medical sciences. My name is Meryem Talbo, and I’m a PhD candidate and a podcast team member of the McGill Journal of Medicine. Please welcome my colleague, Nadia Blostein.

0:56 Nadia Blostein (NB): Hello, my name is Nadia Blostein and I am currently a Master’s student in Neuroscience at McGill University. As a member of the McGill Journal of Medicine Podcast team, I had the pleasure to interview medical physicist, Dr. Shirin A. Enger, about her group’s work in the automation of brachytherapy workflows. Dr. Shirin A. Enger is currently interim co-director at the McGill University Medical Physics Unit. She is also Canada Research Chair in Medical Physicals, an Associate Professor at the Gerald Bronfman Department of Oncology, an associate member at the departments of Physics and Biomedical Engineering at McGill University, and the Research Director of Translational Physics and Radiation Biology at the Lady Davis Institute and Segal Cancer Center of the Jewish General Hospital.

1:46: Today, you may expect us to touch onto the interdisciplinary and translational nature of medical physics research, the workflow of radiotherapy, which is a form of cancer treatment that uses ionizing radiation in order to kill cancerous cells that are usually in tumors, and we conclude with a list of open-source initiatives that Dr Enger’s group has been a part of, including the McMed Hacks Summer Workshop series on Machine Learning and Medical Imaging, spearheaded by Dr. Enger’s students Lucas Weishaupt and Yujing Zou. Links to anything discussed in this episode will be included in the show notes, and I hope that you get as much out of this conversation as I did.

2:45 NB: So I guess maybe we begin with a brief overview of what a medical physicist actually is, and what they do and what the difference between a clinical medical physicist and a research physicist would be.

Dr. Shirin A. Enger (SE): So thank you very much, Nadia, for this introduction and thank you for inviting me to talk to you about my research. So, you can be a clinical medical physicist and a clinical medical physicist works to ensure that radiotherapy treatments are safe. So that all the machines are delivering the dose that we claim that they are delivering, so they do very proper and detailed quality assurance of all the machines at the clinic in the radiotherapy departments. They do dosimetries and maintain a safe treatment to the patients.

3:32 NB: Just a quick question about radiotherapy specifically, is that only ever used within the context of cancer treatment or is it used more broadly in a medical context?

SE: So radiotherapy is used to treat cancer patients, and about 50% of all cancer patients receive radiotherapy during the course of their treatment. So, all our patients who have cancer, they can get surgery or chemotherapy or radiotherapy as monotherapies, but all of these different treatment modalities can be combined as well sometimes to ensure that the patient is cured. But radiotherapy is also used as a palliation from the disease, sometimes if the disease burden is too high, just to relieve the patient of some of the pain. So mainly, radiotherapy is used to treat cancer.

4:23 NB: So then in a clinical setting, what a medical physicist does is very much surrounding the machinery as well as the dosimetry of these treatments?

SE: Yes.

4:31 NB: Moving into the scientific aspect, what sort of research do medical physicists engage in?

SE: So medical physicists in general, both clinical and research based ones, are very multidisciplinary as scientists and it is a very multidisciplinary profession. So we use physics in medicine, however, we have to also know a lot about software development, scientific computing, other parts of say, biology, anatomy, oncology and so on and so forth. So we have to have a very broad umbrella and we have many different kinds of scientists that we collaborate with but also we need to know a little bit about everything to be able to do our profession in a good way. For research medical physicists, in most of the academic centers, the medical physics academic unit usually is located at the hospital, so, also McGill. And at McGill, we have several teaching hospitals, one of them is McGill University Health Centre, the other one is Jewish General Hospital, and we provide radiotherapy to our patients in both of these clinics. But also, we researchers are very close to the end-users, both patients and clinicians, so we can see daily what is lacking from current treatments. What are the limitations? And sometimes we come up with totally new devices, softwares… so we work both with software and hardware… And sometimes, we modify the current modality, whatever it is. If it’s an imaging modality or treatment modality, we modify it such that we answer the limitation. So we constantly collaborate and communicate with clinicians to understand the needs. There are unmet needs, obviously, because nothing is perfect, and continuously there are situations where an equipment needs to be further developed for either a totally new application or for its current application it’s very limited. When we sit in this environment, it’s very rich: we have many other professionals, such as radiation oncologists, nurses, therapists, dosimetrists, many other professions as well. So, when we communicate with them first of all our students learn how to collaborate and communicate in a very multidisciplinary environment and also it’s a very rewarding research area because if you develop something good, it actually will be used in the clinic, you know? So it’s not just something that you develop and put on the shelf in the lab and nobody else will use it. If you develop a piece of software and you do it well, and actually answer this clinical unmet need, your software will be used. So our research is, at least mine – I shouldn’t talk for my colleagues – is very translational, and I chose to do so from my younger years working with basic science; I’ve moved to translational science because for me, it’s more meaningful.

7:36 NB: And so, in your situation then, would you consider yourself both to be a clinical medical physicist and research medical physicist? Or you’re still very focused on mostly the research aspect, and then in a clinical setting most of what you do is observational?

SE: Yes, I myself never do clinical work. But I always ensure that I have clinical physicists in my team as collaborators and co-supervisors of my students as well as people from other medical professions such as radiation oncologists, medical oncologists, pathologists… Any kind of clinical profession that we are working with – I make sure to include them in my research team as a collaborator, co-supervisor of my students so that we don’t end up developing something that is useless. We have to have this constant communication to make sure that we are actually answering for this need that exists, not just theoretical, very fancy equipment that we develop and then we go, when it’s finished, to the clinicians and then they tell us, oh, this is not it. Which I have done, you know, in the past, when I was a PhD student, so I want to avoid this. So currently, almost all of my PhD and Master’s students are co-supervised by a clinician. Sometimes, I lack certain practical knowledge, because they work with these machines day-in and day-out, and also even when we develop software, how will this software be used in a clinical environment? Will this work with the clinical workflow? Also it’s sometimes fulfilling to work with the person that is doing things rather than, you know, “I do more research”. There are different roles, but we definitely communicate and we are very lucky here at McGill having this kind of opportunity to actually work with clinicians. So as a scientist, I think it is a great opportunity to use and we gladly involve clinicians in our research.

9:28 NB: And how many students do you currently have?

SE: I think we are 24.

9:35 NB: Huge research group! While we’re still on the topic of what a medical physicist does, what is the pipeline of the involvement of all these different people in, say, giving somebody a radiation treatment?

SE: So all of these professions are important in different parts of the patient’s treatment. So, the patient obviously sees the radiation oncologist first, who is a specialist on certain tumour types, and then this radiation oncologist prescribes a certain radiation protocol for this disease site and they prescribe dose scheduling. So the patient comes in and imaging is done throughout the course of the treatment, and the technologist does all the imaging with the patient and then these images are used later by the physicist and later dosimetrist to do patient-specific dosimetry. So we use very advanced treatment planning systems to plan on these images before the treatment is delivered. We plan the maximum dose to the tumour, while sparing organs at risk and whatnot, and this goes to QA (quality assurance), then to the oncologist again and if this is OK, then the patient is treated. And those who treat the patients on the machines and running the machines, those are radiation technologists. So, there are different people who are specialized in different areas during the course of the treatment, but any kind of new techniques and new technology, the medical physicist is involved. And also, medical physicists’ responsibility is that the machine is working and delivers the dose that we have prescribed to the patient. So they have measured, in big water tanks, ahead of the treatment so all the machines are quality-assured.

11:18 NB: Moving into the more specifics: so there are different subfields in medical physics, you specifically are in radiation oncology medical physics. What are the other subfields and how are they different from the radiation and oncology subfields of medical physics?

SE: Traditionally there have been two streams: one is radiation therapy – people working with delivering dose to the patient – and then people work in the nuclear medicine department with imaging. So you can be an imaging physicist or also, in the radiotherapy department, a diagnostic radiology physicist. Thus, there are physicists in charge of all the imaging equipment in radiotherapy in clinics, and those who only work in nuclear medicine with nuclear medicine-based imaging. I don’t know if you know the distinction, but in diagnostic radiology, we have computed tomography (CT) and MRI machines and these give us anatomical information. We take images of the anatomy of the patient, while the nuclear medicine department has positron emission tomography (PET) and SPECT, two molecular imaging techniques. Thus, one type of physicist uses magnetic resonance (MRI) or X-ray (CT) based imaging, while in nuclear medicine department we inject radionuclides into the patient and use say positron emitters or other photon emitters in SPECT (single photon emission computed tomography) to image the patient. So these are different techniques, and thereby different specialists are needed for these departments. More and more, theranostics are becoming popular nowadays. The radionuclides that you use for imaging can also be used for therapy. So when imaging a certain tumor type with nuclear medicine based imaging modalities, some research groups also try to combine therapy with imaging. These kinds of techniques are becoming more and more popular in both research and preclinical studies.

13:28 NB: I’ve heard that they’ve been trying to do that also with ultrasound imaging. I don’t know what kind of subfield of medical physics ultrasound imaging would fit under, but I heard that they can inject people with certain particles and then use imaging to have the particles release some sort of treatment molecule in a very spatially-specific fashion. Would that fit under the field of nuclear medicine?

SE: No, that’s still done in any sort of imaging department. In nuclear medicine, we have radioactive isotopes, in very small concentrations. For imaging, if we want to use this for therapy, we have to first target it only to the tumor but also to the activity levels that need to be increased.

14:11 NB: So then in a diagnostic setting, the point of these nuclear medicine imaging techniques is that you have a specific molecule, and you’re trying to see where in the brain it’s present and in what quantity, whereas if you do it in a more treatment setting, it’s going to be to have that molecule delivered to a specific tumor.

SE: Yes, and this is a very big research area right now and if we can make sure to deliver, with whatever technology we use – there are many ways of delivering these compounds to the tumor – I think this will be the area to focus on because we are concentrating these targets inside the tumor and sparing health tissues. It’s a promising area but it’s new – dosimetry is not as developed as in the traditional or conventional therapies – so there is a lot to be done there, and it’s a very active and interesting area of research right now.

15:07 NB: So this leads into the next topic, because you were talking about treatments that were very tumor specific and spare other tissues, so would you like to explain what brachytherapy is and how it’s different from more traditional approaches to tumor treatment?

SE: So brachytherapy, if we’re talking about traditions, is the oldest radiotherapy treatment because it was utilized directly after Marie Curie discovered radium. So it is the first radiotherapy treatment! It has developed and evolved during the years. So, to just give a little background information, radiotherapy can be delivered through external beam radiation where you have a linear accelerator, which is a big machine, delivering radiation from beams from outside in. So you have a beam that is coming from this machine and targeted towards the tumor inside the patient and delivering the dose. Although we rotate this machine around the patient to make sure to concentrate the beam to a specific area or to this target volume (which is the tumor), and although most of the dose will be delivered inside the tumor, we still have to go through some of the healthy tissues to treat the tumor. On the other hand, in brachytherapy, we place the radionuclide inside or very close to the tumor, giving radiation from inside-out, so we can give a very, very high dose to the tumor and spare healthy tissues as much as possible. These little seeds that we have in brachytherapy, they are as small as a rice corn, and they have to be delivered to the tumor in one way or the other. Usually brachytherapy treatments are used for tumor types close to a body cavity, such as gynecological cancer, rectal cancer, lung cancer, esophagus, and so on. But also it is used for prostate cancer because it’s very closely located to the prostate, so we can implant very thin needles, sometimes up to 21 needles inside the prostate. And then this radioactive source itself, it is connected to a metal guidewire that sits inside a machine called an after-loader. It’s a robot, you could say. And this is high-dose brachytherapy. So when the patient comes in, the patient is imaged, and then something called an applicator is placed inside the patient, and brachytherapy is used for head and neck cancers, because it gives a very high dose and we have a very high cure rate without doing any operations on the face and so on. For the quality of life of the patients, it’s very important to maintain the organs as much as possible.

17:45 NB: The needles are small enough such that they can be inserted into somebody’s head, it won’t really leave a scar?

SE: If it’s this close, we go through the nasopharynx. So we place these devices called applicators inside, say, the nasopharynx (if I have an esophageal tumor), to help guide the radiation source, and usually applicators have places for some small needles. So the needs are already incorporated in this device. And for gynecological cancers and rectal cancers, we place an applicator that holds the organ in place, and contains certain holes for needles that can guide the source to the treatment area. And when it is interstitial cases, we place the needle inside the organ itself. In the case of prostate, these needles are very thin, so they are placed inside the prostate and then the patient is imaged with this applicator in place; it sends the image to a treatment planning system where the treatment is planned. We make sure to give the prescribed dose to the tumor, so there are mathematical optimizations involved here. So you have to make sure to optimize the dose to the tumor, sparing healthy tissues, there are some penalties put on those delivering too much to the healthy tissues because we want to cure the patient but at the same time maintain a good quality of life for the patient. So then something called the transfer tube is connected to these needles or applicator, and the treatment plan that is made is sent to this robot, and this robot sends the source into the needles one by one, places the source at different positions, irradiating for different amounts of time according to the treatment plan that we make, and then retracts the source. And this takes minutes, a couple of minutes, all of this, it’s very fast, it goes in the source, irradiates, goes back, comes to the next needle. If we have several needles in the applicator or implanted in the patient for prostate, for example, the robot does everything and during the time the robot is sending the source in and out, the medical physicists and radiation oncologists need to leave the room, for radiation safety purposes. And then the source is retracted to this robot inside a safe and the needles are removed, so the patient does not contain any radioactivity and the source goes back. The patient might come back. Sometimes, this is a schedule that we follow. So sometimes, the patient has to come back two times, three times and so on. It’s a fractionated schedule, but usually, in brachytherapy, this fractionation schedule is much less than for external beam. In external beam, sometimes the patient has to come back 50 times or so, depending on the protocol we are following, but in brachytherapy, we give a very high dose during a very short period of time and the patient comes back a couple of times sometimes – up to 5, 6 times, for certain tumor types – but not as often as for external beam.

20:50 NB: And I guess I was wondering – when you were talking about the needles – how do you pick the number of needles? Because intuitively you would assume that, say you have 100 needles, you get a lot more, I guess, spatial specificity? But then it might be longer to place? So then is there a trade off, you know, you want to limit it to be less invasive, but then you want to get a good enough spatial resolution… So is that part automated as well?

SE: Unfortunately not, and that is actually an active project in my lab. So right now the tumor where most needles are placed is the prostate, and we can sometimes place up to 21 needles. And these needles have an outer diameter of 2 millimeters, so they are pretty thin. But still, oncologists place these needles without any optimization, by just looking at the images. And these are done under ultrasound guidance, so she sees the organ she’s placing the needles in, but this is not at all optimized. So right now, I have an active project in my lab. We are trying to first optimize the needle placement, but also to automate it. So this is a very good question, needle placement currently is not optimized at all. So the dose delivered through this needle are optimized, so already, we get a non-optimal situation where the oncologist has placed the needles and then it’s our job to make sure that wells are positioned and timed so that the source gives us an optimal dose to the tumor while saving organs at risk. So I think by optimizing needle placements to begin with, we will ensure, for example, that fewer needles are placed. Because right now all 21 needles are placed, sometimes we don’t activate all of them. So when we make the plan, we realize, OK maybe 3 of them or 4 of them are just not needed. So that is why it’s good to be multidisciplinary and work together to solve such a problem as you mentioned.

22:50 NB: So the variables are how many needles, where do you place them, and then also in each needle, how long and how much radiation do you deliver. So it’s four variables that could be automated, hopefully in the future.

SE: Yes so the placement, well, where this radiation source should be placed inside one needle and how long in each of the positions, because time plays a very big role here. This we have already automated. Our optimization and techniques and tools in this are developed by my lab because the commercial ones are not very good, in optimization techniques, so we have already published and it is actually a software engineer who developed this. Because you know, students in my lab come from all sciences… from physics, mathematics, but also engineering backgrounds. And some of them are interested in applying their degree, might be in software engineering, electrical engineering, whatnot, to medicine. So medical physics is a very good area to apply because we have many projects which are clinical yet require knowledge from another field, such as software engineering. So the student that I had, Majd Antaki, he did develop, he just graduated, and got a job at a medical company, working on exactly the same thing he did in my lab but getting paid much more (laughs). So he did develop these optimization algorithms for the well time and position optimization, so now I have another student that is optimizing and automating the needle placement part of it.

24:31 NB: So it’s all coming together, all these projects. I was also wondering, are patients usually anesthetized when they’re undergoing brachytherapy?

SE: Not for all treatment sites, but obviously when interstitial needles are used, so when you are poking the patient with a needle, then they are anesthetized. And it’s also different between different hospitals. So for intracavitary cases where we place an applicator inside the rectum, or for some gynecological cases, you know, anesthesia is needed. So anesthetized, yes, but for certain tumor types.

NB: And I’m guessing that anesthetics don’t really have an impact on this sort of treatment?

SE: No, they don’t. But many times in the clinic the bottleneck is actually waiting for somebody that can come and deliver anesthesia to the patients, but also if we can shorten this time for the patient experience, it is much better. So automation – any kind of automation and optimization – obviously ensures that the patient is treated during a much shorter timeline, and less invasive. So the patient experience is improved.

25:38 NB: So speaking of saving time, would you like to talk about the RapidBrachyDL algorithm that was developed by your group that radically saves time in the process of optimizing dosimetry?

SE: Yes, so RapidBrachy DL is predicting dose, to just put it simply. Because in brachytherapy, for example, the clinical treatment planning systems assume that the human body is an infinite water sphere. And they use precalculated dose kernels in the dosimetry because they want things to go fast. In the clinic you want everything to go fast – especially if the patient is anesthetized and whatnot – but doing so, assuming that the patient is a big sphere of water, is obviously very approximative, because obviously we are composed of water, but not totally. For certain tumor types, this approximation is less accurate. For some of them it is more accurate, for some of them it is less accurate. So the best way to calculate dose in a very heterogeneous system such as a human body is to use Monte Carlo simulations to track radiation in each tissue type and calculate dose accurately. Monte Carlo simulations are the most correct method to calculate dose. However, they are also the most time consuming. So accuracy and time – they go hand in hand. So in my lab we’ve developed a treatment planning system called RapidBrachyMCTPS. It is a totally Monte Carlo-based treatment planning system and is as complete as a clinical treatment planning system from using patient imaging to doing optimization… everything. You can do a complete treatment plan in our software, however, our dose calculation engine is Monte Carlo based, which is very accurate, but very slow, so this can never be used in a clinical situation because it’s not Health Canada approved, so it can’t be used in a clinical situation, but it can be used to benchmark a clinical plan. When we suspect that, oh, here we could have a big discrepancy between prescribed and delivered dose. With a clinical plan, we can do one with mine to see, how much off are we? But if mine takes one day to calculate, obviously, that’s not an option. So then I was thinking if it’s possible to predict dose? To actually train an algorithm that can predict radiation interaction with the human body. Dose is energy deposited per mass. So then we developed RapidBrachyDL that actually can predict doses very accurately. So we have our Monte Carlo based dose map as a ground truth, and trained algorithm with the Monte Carlo based dose maps in patient and the patient geometry, and the results are very good. We trained the model only for prostate and then transferred the study for cervix cases. Right now I have a PhD student who is training the algorithm for all other tumor types that we treat with brachytherapy to make sure that RapidBrachyDL becomes a general purpose dose prediction model so that we can have that also in our treatment planning system. Dose researchers and clinicians who use our system and are interested in developing new technology can use the Monte Carlo engine, obviously, but those who are just doing routine dosimetry then they can use RapidBrachyDL, which calculates dose in seconds, compared to minutes and hours.

28:51 NB: And so a fundamental difference is the Monte Carlo simulation uses all these different mathematical models and physical properties inside a specific patient's organs and tissue types and that’s why it takes so long. On the other hand, one can use deep learning to make rapid predictions. Deep learning, which is a technique that uses neural network architectures, is a subcategory of machine learning which in itself is a subcategory of artificial intelligence. The idea is that to build your deep learning model, you need to start off with a dataset, where for each data point you have numerical input features and numerical output features. For instance, if your dataset is comprised of 20 CT scans from 20 different patients, the input features for the CT scan of 1 data point, a patient, could be a numerical intensity value at each voxel, which is a 3D pixel, in that patient’s CT scan, and the output could be a numerical voxel wise dosimetry value, which has already been obtained through Monte Carlo simulation. Once you curate your ground truth dataset, you can build your model by training a computer to make connections between different input values and output values. This is a training phase and it requires a lot of data and computational resources. Once the model has learned how to make connections between inputs and outputs, you can feed it patient-specific inputs and it will rapidly output a voxel-wise dosimetry prediction, without needing to perform a bunch of complex calculations on the spot, which is why RapidBrachyDL is a more efficient tool to use in a clinical setting than Monte-Carlo simulation. I was curious about the fact that your algorithm was trained on voxel-wise data, right?

SE: Both yes and no. So yes, we use voxel-based dosimetry. We calculate dose per voxel, however, these voxels are contained within an organ. So we calculate organ-based voxelization. So it is dose per voxel but you have millions and millions of voxels. If it’s a prostate cancer patient, we have the entire abdomen. And so on. It is voxel-based but it also takes into account the entire patient geometry.

NB: So basically you have a certain amount of input features per subject and you train your algorithm for each subject, it knows how to link these input features with a specific output which is the dosimetry. So a traditional machine learning problem is the curse of dimensionality, where you have more input features sthan you have subjects. Was this a problem for rapid brachy DL and if so, how was that problem circumvented?

SE: Yes, so the first thing that we did was to decide what is most important because voxels further away from the source for example, we cropped. In the Monte Carlo dose map, we kept the patient as is, because for Monte Carlo simulations it’s important the radiation that scatters inside the patient also delivers dose, so we kept that. But for the deep learning based one, firstly we cropped the patient geometry so that it’s much smaller, removing all the voxels that don’t contribute or play a role in this case. So first of all we have to think that this algorithm is predicting radiation interaction with matter. So it’s only going to predict dose correctly for the organs that we’ve trained it on. So right now it predicts very well for the abdomen area, and it really doesn’t matter if it’s a male or female because the soft tissues are the same. The interaction probability of radiation with these organs or tissue types are the same because it’s the same soft tissues, bone and so and so forth. Then we are moving towards the breast, for example, and we have other problems, because in the breast, on one side we have the tissues and lungs, air, and so on, and ribs are much closer, but on the other side we have air. We will have a problem with reproducibility if we use this algorithm, say, to predict dose in breast cancer patients, because we haven’t trained the machine to predict dose in air for example. So right now, we are doing a lot of training on breast cancer patients and also on applicators with different materials who are not tissue-equivalent, metallic ones and hard plastic ones, and whatnot.

NB: Just to put things in numbers, this algorithm was trained and validated on a total of 47 patients with prostate cancer. During the testing phase, rapid brachy DL predicted dosimetry in both prostate and cervical cancer patients with an extremely high accuracy relative to the ground truth, which was calculated via Monte Carlo simulation. This showcases how robustly this algorithm generalized to unseen cervix data. Finally, the deep learning algorithm reduced dosimetry prediction calculations from 8.9 minutes with Monte Carlo simulation to a very brief 1.1 seconds. Moving forward, throughout the brachytherapy process, there is a lot of image segmentation that needs to be done so just for people who don’t know, segmentation is when you have a 3D image and you’re looking for a specific organ, you have to label it first in that timage before doing other computations on it. So, is that an avenue that your lab is exploring in terms of automatting segmentation?

34:06 SE: So we are also working on segmentation because in the treatment planning workflow, as you said, we need to have segmented images. We need to know where the tumor is and where the organs at risk are because we calculate dose in a volume. So this volume needs to be defined. So we voxelize each of the organs that have been segmented by a radiation oncologist. Obviously we would like to remove this human annotator. So right now, we are trying to automate the entire treatment planning process and this means segmentation. In the brachytherapy also we have to digitize implanted needles and applicators because these are implanted inside the patient. We take the image, but then we need to know where is this applicator, where are the needles, to be able to do the treatment plan, to place this radiation source inside this foreign object that we placed inside a patient. So we are also digitizing the applicator and needles with help of AI now, so we are doing that as well. And the dose calculations with rapid brachy DL we are trying to automate it, but also the optimization as well. So my dream is that we just upload the images and a couple of seconds later comes the treatment plan. So everything should be automated. And we are working on all of these parts, one by one to then later put them together in one software.

35:28 NB: Hence a 24-person lab to elegantly merge all these moving pieces. To wrap up, are there any open science or education initiatives that your group has been a part of? Personally I’ve heard about it through McMed Hacks last summer, which was a workshop hosted by you and your students about the use of machine learning in medical imaging. Are there any other initiatives and is McMed Hacks happening again this summer?

SE: Yes it will go this summer again, we are working on the syllabus right now, and soon we’ll go into advertising. So definitely we will have the workshops back this summer. But also, on the treatment planning system that we are developing called rapid brachy MCTPS, it is open source, so and right now it is all Monte Carlo based. Well, the calculation engine is Monte Carlo based, but the rest has been developed by my students, and amongst others, one medical student and some computer science students, and medical physics students… and so, the students developing this have been from all over the place. It’s all departments at McGill, jointly (laughs), helped to develop this. That is open source and we helped users in Europe, Australia, North America, all over the place. I think the only continent we don’t have users is Africa, and I hope we will have African users as well. This is a very big open source initiative and this is the only Monte Carlo based treatment planning system worldwide, so there is none actually, none commercial, none by other labs. And making this fully automated, I think, will be also of great benefit for everybody that is using it right now because it will be much more faster. In addition to these software initiatives, we do also develop radiation detectors and dosimeters. Also, I have a radiobiology lab where we are looking at radiation interactions with different tumor types to investigate which radiation quality is better than the other, because for a person coming from outside, you know, they treat radiation as one unknown parameter that they know of that can be given and it treats certain tumors, and it kills something. But depending on the type of radiation and the energy of that type of radiation, it does cause different amounts of biological damage. So we are looking at different radiation qualities that we have in our hospitals but also up and coming and emerging radiation qualities to investigate which type of radiation is better suited for which tumor types. I do have a wet lab as well where we investigate this, and at the detector level we actually build devices but also brachytherapy radiation delivery systems where we put shields inside applicators and needles that I talked about that rotate during the treatment, and deliver dose to the tumor while shielding organs at risk, because brachytherapy seed gives a spherical dose distribution, so if you have a radionuclide that emits radiation, it emits radiation isotropically, so it emits in all directions and it is seldom that we have a perfectly spherical tumor where we can place this seed inside and it is just traced the tumor. Normally the tumors are really irregularly-shaped, so when we put the source inside, obviously, treats the tumor but also some of the dose spills over to organs at risk so although brachytherapy is much more forgiving in sparing healthy tissues, it still delivers some dose to the healthy tissues and our patients still have some side effects. So by placing shields inside these needles and applicators we have an emission window, We have a groove channels for the needles inside the shield and this shield has an emission window and these shields are connected to a stepping motor as a rotating panel who rotate during the treatment, make sure to always give the radiation to the tumor while shielding nearby healthy tissues. We developed some devices as well. So it’s a little bit of everything.

NB: Wow. That is very cool. It seems like it’s such a creative field. Thank you so much, this was really interesting. And I’ll put the links to McMed Hacks, so that people know where to sign up, as well as any relevant papers on projects we’ve discussed in the shownotes. Thank you.

39:39 MT: And that’s all the time we have today. Thank you for joining us on another episode of the McGill Journal of Medicine, the MedTalks series. This podcast was edited and produced by MJM’s podcast team members, in alphabetical order: Nadia Blostein, Sarah Grech, Neeti Jain, Predrag Jovanovic, Esther Kang, Katherine Lan, Tom Lee, Laura Meng, Amanda Sears, Meryem Talbo and Susan Wang. Feel free to reach out to us on Twitter or Instagram @mcgilljmed or by email. Would love to have your feedback and don’t forget to join us again for our next episode.



Creative Commons License
This work is licensed under a Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International License.