The human brain is an extraordinarily complex organ responsible for our thoughts, memory, breathing and so much more. Its intricate networks, made up of billions of neurons working together, make our actions possible. Neurodegenerative diseases like Alzheimer’s and Parkinson’s disease can impact the patterns of these cells’ activity, and being able to map these networks and the cellular activity within them can inform potential treatments for these conditions.

Scientists have recently developed powerful microscopy systems and molecular techniques that create three-dimensional images of cells to study brain function and activity in detail. However, these images are exceptionally large and complex (with trillions of pixels), making it very difficult to detect changes in network activity patterns.

To address current gaps, a group of researchers in the Hurvitz Brain Sciences Research Program, in collaboration with teams in the United States and Europe, developed the AI-based Cartography of Ensembles (ACE) pipeline, a software that identifies patterns of brain cell activity in large volumes of brain data. A study describing the ACE pipeline architecture and its application in complex neuroscience problems was recently published in Nature Methods. ACE was designed using cutting-edge deep learning algorithms and trained on more than 30,000 3D images curated from microscopy images.

“ACE is capable of analyzing a wide variety of microscopy images, meaning researchers can use the pipeline to gain new insights into how specific populations of cells in different regions of the brain respond to disease,” says Ahmadreza Attarpour, the first author of the study and PhD candidate at SRI and the University of Toronto. “ACE goes beyond traditional methods relying on brain maps that divide the brain into pre-defined regions based on their coarse structural differences.”

Our novel pipeline acts as a detective, pointing out cell activity and patterns that would be otherwise difficult for even highly trained professionals to identify.”

The tool has the potential to accelerate discoveries in neuroscience because of its ability to help researchers accurately identify patterns of activity in specific cell groups and networks within every region of the brain. Researchers can monitor these patterns to better understand how neurological diseases affect brain activity and how treatments may normalize these activity patterns.

“Using ACE, scientists can evaluate the effects of experimental drugs on a particular population of cells across the brain or identify novel targets for neuromodulation therapies,” adds Dr. Maged Goubran, scientist in the Hurvitz Brain Sciences Research Program and Physical Sciences Platform and co-senior investigator of the study.

“ACE provides a powerful tool for mapping brain function and circuitry, paving the way for breakthroughs in neuroscience research and, ultimately, improved patient outcomes,” explained Dr. Bojana Stefanovic, senior scientist and director of the Physical Sciences Platform at SRI and co-senior investigator of the study.

The post AI in Action: Monitoring cellular brain networks appeared first on Your Health Matters.

Tony Xu, PhD student in the Department of Medical Biophysics at the University of Toronto, was awarded the 2024 Google PhD Fellowship in Health & Bioscience. The Google PhD Fellowship is a highly competitive program that recognizes outstanding graduate students doing exceptional and innovative research in areas relevant to computer science and related fields, providing mentorship and funding to advance their work. Tony is working at Sunnybrook Research Institute (SRI) with Dr. Anne Martel, senior scientist in the Odette Cancer Research Program and Dr. Maged Goubran, scientist in the Hurvitz Brain Sciences Research Program. His research at SRI focuses on using self-supervised learning (SSL) to tackle key challenges associated with analyzing large and multidimensional medical images.

Medical images, like whole-slide histopathological images and light-sheet fluorescence microscopy images, have incredibly high resolution to resolve microscopic, cellular detail so clinicians can gain insights to discover complex biomarkers for illness. However, given the large size and density of information displayed in these images, manual analysis can be both time-consuming and difficult for the human eye.

“My research looks to broadly improve on traditional deep-learning methods to speed up the analysis of medical images used by other researchers, and to bring deep-learning closer to clinical application,” explains Tony. “My research has the potential to help researchers analyze data faster and with less-effort, increasing foundational research and helping it reach clinical applications.”

Deep learning is an AI method that recognizes complex data to produce accurate predictions and insights, inspired by the way the human brain would. Many traditional deep learning methods are heavily dependent on expert annotations, or labeling individual elements in data sets to help machines understand the contents, which can take up to weeks to create for just a single image. Tony’s research with SSL works by training deep learning models on unannotated images to teach it to recognize and predict missing data, significantly decreasing the need for expert input.

“Data in the medical domain is not only becoming increasingly available, but also increasingly dense,” adds Tony. “New methods that are able to learn to extract information and patterns from raw data, can play an invaluable role in creating powerful, generalizable models that can be trained just once and applied to a multitude of clinical tasks.”

AI research like Tony’s has the potential to dramatically improve the workflow of researchers and clinicians alike, impacting patient outcomes by assisting clinicians with analyzing images and diagnosing disease and promoting more fundamental health care research.

The post AI in Action: Improving automated image analysis appeared first on Your Health Matters.

Prostate cancer is the most common cancer in Canadian men, with more than 27,000 Canadians being diagnosed with the disease each year. In many cases, prostate cancer develops slowly and can be successfully removed or managed before it spreads to other parts of the body. However, like most types of cancers, there is a risk of prostate cancer to spread or recur after removal or treatment. Traditionally pathologists and clinicians determine cancer recurrence by manually analyzing different images or biosamples.

Matthew McNeil is a senior PhD student working in senior scientist Dr. Anne Martel’s lab at Sunnybrook Research Institute (SRI) where he is developing tools to support the automatic prediction of cancer recurrence. This fall, Matthew was the winner of the Leopard Challenge, a global AI competition. Specifically, the challenge focused on yielding deep learning solutions to predict the time to biochemical recurrence of prostate cancer from H&E-stained histopathological tissue sections.

Matthew is developing AI models that quickly detect features, like Gleason patterns, on slides that pathologists typically use to determine prognosis. The Gleason classification system is used to grade prostate cancer. The scale looks at how abnormal glands in the prostate look and helps determine how likely the cancer is to grow and spread.

“By having these types of patterns highlighted automatically, pathologists will be able to analyze slides more quickly and effectively,” explains Matthew. “My model also has the potential to provide clinicians and patients with a better understanding of the cancer’s risk.”

The model is capable of quickly generating risk scores for patients with prostate cancer. This has the potential to advance patient care and outcomes by supporting more personalized treatment plans as generating these scores can help both clinicians and patients make more informed care decisions.

Matthew is hoping to apply the model he developed during the Leopard Challenge to other datasets for different types of cancers and see his work expand to clinical settings.

The post AI in Action: Predicting prostate cancer risk appeared first on Your Health Matters.

EDs across Canada continue to face pressures including staffing challenges and more complex patients who are more likely to need hospital admission. As Canada’s first and largest trauma centre, Sunnybrook’s ED provides care to some of the sickest, most critically injured patients in Ontario, meaning patients with less urgent medical needs sometimes take longer to be seen. For many patients, visits to the ED may cause anxiety, and not knowing when you might see a doctor may heighten these feelings of stress. Sunnybrook is improving the patient experience in its ED by using AI to share predicted wait times with patients during their visit.

In 2022, Sunnybrook launched a wait time prediction tool on digital screens in the ED and on its website. Developed by Sunnybrook’s Decision Support Team, the display is powered by a machine learning algorithm that handles categorical data and shows the predicted maximum time the majority of patients will wait until they see a physician. Every 15 minutes, a snapshot of the current state of the emergency department is captured. This snapshot assesses information like the number of patients waiting, the day of the week, time of day, number of highly acute patients, wait time patterns, and other relevant factors.

“Staff and patients were surveyed about their experience with the wait time predictor tool, and patients and families appreciated they were provided realistic expectations about wait times,” explains Erin Scholl, Director, Corporate Performance and Business Analytics and lead for the initiative. “With information from the wait time tool, they could keep caregivers informed, or plan ahead for food or parking arrangements.” The prediction tool is also available to the public on Sunnybrook.ca/wait, providing patients with information about wait times before they even arrive.

The tool has also improved the working environment for ED nurses, doctors and staff, equipping them with informed answers to questions like “how much longer?” and ultimately reducing tense interactions with patients.

“The wait time prediction tool provides our patients with transparency from the moment they begin their care at Sunnybrook,” says Dr. Justin Hall, chief of the Department of Emergency Medicine and program lead of the Virtual Emergency Department. “By providing patients with this information, it encourages them to assess the reason for their visit and consider alternative forms of care, such as an appointment with the Virtual ED.”

Dr. Hall adds, “it’s important to point out that anyone who thinks they may need emergency care should not hesitate to come to the ED. All patients are important and will be seen.”

Predicted wait times for patients can vary based on the status of their condition, but as more insights from this initiative are gained, there’s a hope the tool will be used to generate patient-specific wait time predictions based on the information obtained at ED triage. The use of AI and machine learning provide a promising future for operations in the ED, as insights learned from this project and other AI tools will help guide staffing recommendations and resource allocation.

The post AI in Action: Empowering patients in the emergency department appeared first on Your Health Matters.

Breast cancer is one of the most common types of cancer affecting women in Canada. About 1 in 8 women develop breast cancer in their lifetime. Starting in the cells of the breast, tumours can infiltrate nearby tissue or spread to other parts of the body. Breast cancer can be treated with surgery, drug therapy, and radiation therapy, but not all breast cancers are the same. Since every patient will respond to these treatments differently, predicting a tumour’s response to therapy can improve patient outcomes.

Dr. William Tran, radiotherapist and senior scientist in the Odette Cancer Program, has teamed up with Dr. Fang-I Lu, a breast pathologist at Sunnybrook and associate professor at the University of Toronto and together, they are developing AI technology to improve radiation therapy treatment for breast cancer patients.

The team is mapping the tumour immune microenvironment, the complex ecosystem of cells surrounding the breast’s tumour. Since the immune system plays a role in clearing tumour cells, the team aims to measure the probability of a tumour’s response to high-dose radiation treatment in women with high-risk breast cancer. The project involves taking thousands of tumour images and complex computational methods to recognize biomarkers associated with the tumour-killing effects of radiation treatment.

Dr. Fang-I Lu and Dr. William Tran analyze an annotated pathology sample.

“Our work will allow doctors to determine which patients will benefit most from radiation therapy,” explains Dr. Lu. “The ability to predict a tumour’s response to certain types of therapy has the potential to support more personalized and effective treatment plans for patients with advanced breast cancer.”

Radiation therapy is a common type of treatment for many cancers. The treatment uses high-energy beams to destroy the genetic material in cancerous cells that control how the cells grow and spread. Radiation therapy can also damage healthy cells in the body, which can cause short-term and long-term side effects like hair loss and fatigue. “Personalizing the treatment plan using AI can help optimize treatment outcomes while minimizing the side effects,” adds Dr. Tran, who is also an associate professor at the University of Toronto. “Our AI-based prediction model will help spare patients who are unlikely to benefit from radiation treatment from the short and long-term side effects associated with exposure to that type of therapy.”

Since 2018, the team has been researching the use of AI and digital pathology to map breast tissue samples and measure their resistance to neoadjuvant chemotherapy treatments. With this newest model currently undergoing planning for early-phase clinical trial testing, Dr. Tran hopes to continue improving personalized cancer treatment planning in the radiation oncology clinic.

The post AI in Action: Supporting precision radiation oncology appeared first on Your Health Matters.

AI is a huge area with many applications. What’s happening at Sunnybrook?

It’s such a vast field, and there is so much going on at Sunnybrook. A lot of the AI research we’ve done at SRI has focused on getting information from pathology and radiology images, like x-rays and scans. We have developed methods to pinpoint disease in an image, and that helps pathologists and radiologists do their jobs better and faster. This also identifies the best treatment for each individual patient. When data from thousands of patients are fed into AI algorithms, that can identify patterns to help treat other people with similar diseases. Historically, finding these patterns would have taken years to do. AI lets us do this more quickly and accurately.

We are also taking advantage of the amazing progress made in developing AI models that are capable of understanding text. When a patient comes into hospital, we are collecting data at many points; the initial conversations at check in, during blood work, imaging and pathology, as well as when treatment is prescribed. All of this data helps us understand each particular patient’s condition, but right now research fellows have to comb over written notes to extract the relevant information. We can harness the recent advances made in AI to carry out this laborious work automatically.

How new is AI at Sunnybrook?

The techniques behind the imaging data collection I mentioned, and the principles of machine-based learning, were developed decades ago. My research team built an AI algorithm capable of analysing breast magnetic resonance images over 15 years ago, and we have continued to explore and develop new techniques ever since.

A few years ago, we gained international recognition after developing the first digital pathology foundation models to exist, meaning we took over one million specimen images from public databases and trained an AI model to understand patterns in what it was seeing. That now helps us figure out what disease a patient may have, or if they are responding to treatment, based on wider patterns. It’s also made it possible for us to develop more accurate AI models more quickly.

AI awareness among the general public and clinicians has increased greatly, and this has led to an eagerness to explore how it can improve patient care.

Some people may be fearful of AI. What do you say to them?

When people think of AI, they often think of ChatGPT and those kinds of applications. At Sunnybrook, we won a Canadian Foundation for Innovation (CFI) grant a few years ago which helped us build a state-of-the-art AI computing platform with enough power to allow us to train and run AI models within Sunnybrook’s secure environment. We are very aware of the importance of protecting each patient’s personal health information, and the platform we use ensures that.

What will AI deliver to patients in the next 10 years?

While much of Sunnybrook’s work in AI still isn’t at the point of directly affecting patients yet, that will hopefully change soon. One shift could be patients seeing faster action. For example, we’ve developed an AI algorithm that can find tiny regions of tumour cells in microscopy images. This could reduce the time patients have to wait for results after surgery. One of my colleagues has developed a model to identify brain bleeds in CT scans, so patients at risk can be seen immediately without having to wait for the scan to be read.

Another area would be personalizing medicine, where using AI could help direct what type of treatment or surgery would be optimal for each patient based on their unique circumstances.

AI may also help us use hospital resources more efficiently, and is a real priority at Sunnybrook. Things like using AI to figure out what bloodwork each patient needs based on their condition, which can cut down on unnecessary tests.

Are there downsides to AI?

It’s important to remember that AI is not magic; the algorithms are only as good as the information we feed into them. If an AI model is only trained to look at one group of patients, for example, there will be a built-in bias if applied to others. We need to train our AI models around a varied patient population that reflect the realities of our communities.

AI data also needs to be continually updated. Imagine that we had an AI model that was trained in 2018 to predict whether or not a patient has pneumonia. Today, in a post-COVID world, that model wouldn’t work anymore.

What about the human touch?

AI won’t replace humans in health care. It’s just a tool. AI can be a powerful aid, but we’ll still need the expertise of our clinical staff to understand how we best use it.

The post AI in Action: The AI landscape at Sunnybrook appeared first on Your Health Matters.

Stroke is a medical emergency resulting in a sudden loss of brain function. Most strokes, also known as ischemic strokes, are caused by an interference with blood flow to the brain, due to a clot, narrowing of a blood vessel or bleeding. In Ontario, there is one new stroke victim every 30 minutes. Dr. Bradley MacIntosh, senior scientist in the Hurvitz Brain Sciences Research Program and member of the Dr. Sandra Black Centre for Brain Resilience & Recovery at Sunnybrook Research Institute (SRI), and his lab are using deep-learning AI tools to support acute stroke imaging and aid in the treatment of patients suffering the deadliest types of stroke.

Although ischemic stroke makes up the majority of cases, an intracerebral hemorrhage (ICH), or brain bleed, is another form of stroke and occurs by a spontaneous artery rupture. Traumatic brain injury can also result in a brain bleed. ICH is the deadliest types of stroke, accounting for one in six acute strokes.

Timely diagnosis and treatment play a critical role in stroke patient outcomes and in order to provide care for patients suffering from a brain bleed, clinicians need to know the size of the bleed and where it is located. The MacIntosh lab has built an AI tool called VIOLA to analyze, detect, and outline locations of hemorrhages visible on computed tomography (CT) scans of patients who have had a stroke. The VIOLA tool searches CT scans and automates these measurements, saving time and potentially improving patient outcomes.

The VIOLA tool can scan CT images and highlight potential brain bleeds in a matter of seconds, assisting clinicians in determining timelier diagnoses and treatments.

“The VIOLA tool can provide clinicians with an accurate analysis and measurement of a brain bleed, enabling faster care and ultimately increasing the likelihood of recovery for stroke patients,” explains Dr. MacIntosh. “The tool also has the potential to create automated and standardized reports, which can assist clinicians in delivering stroke care.”

The VIOLA tool was developed by a collaborative team led by Dr. MacIntosh and researchers at the Computational Radiology and Artificial Intelligence unit at Oslo University Hospital in Norway. The tool is currently undergoing clinical evaluation at Oslo University Hospital, with international collaborators in Sweden and the United States serving as stroke centres for external validation. For the next phase of his research, Dr. MacIntosh hopes to implement the updated version of the VIOLA tool at Sunnybrook. While VIOLA continues to undergo validation, Dr. MacIntosh is using Sunnybrook stroke images to pursue and study new and complimentary AI radiology tools.

“ICH is the deadliest type of stroke and can have devastating, lasting impacts on a survivor’s brain function and physical ability,” adds Dr. MacIntosh. “Stroke researchers are making big progress. There were two landmark clinical trials for ICH in the last year. We created VIOLA with the hope of jumpstarting more ICH research to advance future care for stroke patients.”

The post AI in Action: Locating brain bleeds faster appeared first on Your Health Matters.

Oral cancer is a form of head and neck cancer, affecting more than 5,000 Canadians per year. Progression of the disease can drastically affect the way an individual eats and speaks. Dr. Jesse Chao, scientist in the Odette Cancer Research Program and Canada Research Chair in Precision Cancer Diagnostics and Artificial Intelligence, is developing new AI software to better detect cancerous cells in the mouth and lymph nodes which will support early detection, staging, and prognosis, leading to improved patient outcomes.

Timely diagnosis plays a critical role in the treatment and recovery of cancer patients. The earlier tumours are detected and diagnosed, the more responsive cancerous cells are likely to be to treatments like surgery, chemotherapy and radiation. Unfortunately, the diagnosis of oral cancer and nodal disease can often be a difficult and time-consuming process. Dr. Chao’s research is aiming to close this gap.

“Our team has developed an AI model that examines tissue samples at various magnification levels, imitating the workflow of a pathologist,” explains Dr. Chao. “This significantly decreases the computational load and complexity of integrating tissue and cellular data for precise tumour detection.”

Using samples from over 500 patients with oral cancer, Dr. Chao and his team are training AI software to automate the image analysis of microscopy data. The software first identifies regions in the tissue and pinpoints areas likely to contain tumours, it then confirms the presence of cancerous cells.

Pathologists often have to go through numerous slides and use many different dyes and stains to determine an oral cancer diagnosis. “Automating this process enhances diagnostic quality, advances early detection, supports personalized treatments and ultimately improved patient outcomes,” adds Dr. Chao.

The software can be used by pathologists to help analyze a sample or provide a second opinion. The software also has the opportunity to serve communities outside of urban centres that lack access to quality cancer care and resources, as it can be accessed remotely by physicians and serve as a potential preliminary diagnosis.

Currently Dr. Chao and his lab are applying this model to detect instances of oral cancer, however, with further research Dr. Chao hopes the software will play a role in the diagnosis of other forms of cancer.

The post AI in Action: Improving screening for oral cancer appeared first on Your Health Matters.