Introducing mHealth Into Indigenous Populations

Introducing mHealth into any community is a challenge. With indigenous populations, there is the additional requirement that any solution introduced is respectful of their culture and way of life. Our work focuses on identifying the barriers and facilitators of implementing mHealth solutions in Indigenous populations across low-and middle-income countries- an often overlooked area of research. By highlighting these challenges, we aim to address the significant gap in existing literature.

Our team is also working to create an internet-enabled mobile app for the Ngäbe community, an indigenous group in Costa Rica, to preserve their traditions, including their knowledge of local plants used for medicine, art, and cultural practices. This project aims to bridge technology with cultural preservation, empowering the Ngäbe community to document and transmit their invaluable cultural heritage to future generations.

Enhancing TIRADs Classification with Machine Learning Techniques on Low-Quality images

In rural areas of Costa Rica, access to advanced medical facilities and specialized
diagnostic procedures is often limited, making it challenging to assess thyroid nodules effectively. Over half diagnosed thyroid nodules are actually found to be benign reducing the need to perform invasive treatments like fine-needle aspiration (FNA) in every case. Instead, the TIRADS (Thyroid Imaging Reporting and Data System) classification provides a practical, non-invasive way to monitor these nodules over time. Portable ultrasound machines, such as point of care ultrasound (POCUS) devices, offer a practical and more cost-effective solution for remote areas.

Our aim is to improve the diagnostic accuracy of machine learning models using
suboptimal images from POCUS machines, ensuring that healthcare providers in these
rural settings can more confidently assess and track thyroid nodules. By doing so, we
aim to reduce unnecessary procedures and promote the early detection of thyroid
cancer.

This project is led by Wardah Ijaz from University of Toronto’s Computer Science Department and Hayden John from Queen’s University School of Medicine, under the supervision of Dr. Pascal Tyrrell, with Dr. Charles Yan as the radiology expert

Integrating Advanced AI into Breast Cancer Treatment in Costa Rica

Integrating AI-driven solutions into healthcare systems is inherently complex. In Costa Rica, the need for enhanced breast cancer treatment is pressing, requiring solutions that are both efficient and sensitive to local healthcare dynamics. Our project focuses on revolutionizing breast cancer treatment in Costa Rica by leveraging advanced data science and AI. We aim to assess the efficiency of AI in triaging cases, analyze system deployment, and explore the implications of false negatives. By highlighting these challenges, we intend to address significant gaps in existing healthcare services and literature.

Our team is also working to develop an AI-powered diagnostic tool tailored for Costa Rican healthcare providers. This tool aims to streamline the diagnosis and treatment process, ensuring that patients receive timely and accurate care. This project not only bridges technology with clinical practice but also empowers healthcare professionals with advanced tools to improve patient outcomes.

This project is led by Lolita Aboa an international data science student from University of Toronto, alongside Dr. Jean Carlo, MD from Costa Rica, and Dr. Pascal Tyrrell

The Impact of Artificial Intelligence on Radiology Specialty Preferences Among Canadian Medical Students and Radiology Residents

AI has reshaped healthcare and radiology, transforming diagnostic and prognostic capabilities. A 2018 survey examined how Canadian medical students perceived AI’s role in radiology, but with rapid advancements since then—such as the emergence of tools like ChatGPT—an updated perspective is crucial. In response, a new survey planned for 2023 seeks to capture the evolving viewpoints of Canadian medical students and radiology residents. It will delve into how AI influences their career decisions and educational needs.

Conducted anonymously online, this study aims to provide insights that can inform medical education. It will explore not only awareness and perceptions of AI in radiology but also how these factors shape preferences for specialties and the integration of AI into medical training. Targeting Canadian medical students and radiology residents at various stages of their education and training, the survey aims to support institutions in preparing future healthcare professionals for an AI-driven landscape.

This project is led by Dr. Pascal Tyrrell with co-investigators Dr. Rachel Fleming, Dr. Alexander Bilbily, Dr. Mohammad Damer, Michael Atalla, Donna Mohseni Mofidi, and Eray Yilmaz

Measurement Error in MSK 2D B-Mode US

In the Measurement Error in MSK 2D B-Model US project, the researchers aim to investigate the inter and intra-rater reliability of ultrasound measurements for knee effusion. The study seeks to distinguish between variations in measurements made by the same operator and those made by different operators when assessing pathological fluid in knees.

This research is crucial for identifying instances where differences in knee effusion measurements reflect a true change in fluid levels rather than measurement variability. Currently, the team is in the final stages of data extraction and anticipates completing the manuscript during the summer. 

This research is led by Eray Yilmaz and Khosrow Khodabandehlou

Examining Effects of Data Heterogeneity on Model Performance

Machine learning models in medical imaging are at the core of computer-aided diagnostic systems, but little attention is given to the impact of data-inherent heterogeneity on such models, which presents significant clinical implications. This study explores the impacts of data heterogeneity on a convolutional neural network across a variety of medical imaging datasets, and explores the relationship between heterogeneity and generalizability of the model, providing practical insights for clinical applications of such systems. Current work focuses on testing and analyzing how informed clustering (via feature extraction) compares to that on clusters defined randomly, and establishing a baseline that indicates whether the clustering improves performance, as an additional pre-processing step for practical application.

This project is led by John Valen, alongside Dr. Pascal Tyrrell, and Lucie Yang

HealthDataScienceEd Canada: Forging Inclusive Health Sciences Training and Career Advancement through AI-Enhanced Learning

HealthDataScienceEd Canada aims to bridge gaps in bioinformatics, computational biology, and health data sciences education and career development. It offers a comprehensive learning environment with both online and in-person solutions, leveraging advanced technology like Large Language Models (LLMs). The project promotes collaboration among students, researchers, and professionals while emphasizing equity, diversity, inclusion (EDI), and ethical research practices.

Through a mix of training methods, the initiative accommodates diverse learning preferences and enhances career readiness with skills in EDI, grant writing, and communication. AI-enhanced learning provides personalized, up-to-date content, and AI algorithms facilitate networking based on shared interests. In-person workshops complement online learning for better knowledge exchange.

HealthDataScienceEd Canada aims to empower diverse talent, reinforce Canada’s leadership in BCBHDS, and contribute to the global knowledge economy, fostering an inclusive, supportive, and sustainable future for health and biomedical sciences in Canada.

This project is led by Dr. Pascal Tyrrell and Dhruv Patel, a TCAIREM-sponsored medical student from Queen’s University

NeRF-US : Removing Ultrasound Imaging Artifacts from Neural Radiance Fields in the Wild

This research is centred on developing neural rendering algorithms to address the challenge of 3D reconstruction from 2D ultrasound captures. Current methods in 3D ultrasound reconstruction often rely on manual intervention or predefined assumptions, limiting their applicability in real-world scenarios. The approach adapted NeRF-based techniques specifically for ultrasound imaging, integrating 3D priors through a diffusion model. This method aims to achieve accurate reconstruction in uncontrolled environments, mitigating common ultrasound artifacts. Additionally, a new dataset titled “Ultrasound in the Wild” was curated, featuring real-world ultrasound images of the human knee. This dataset serves as a benchmark for evaluating the method’s performance under realistic conditions and will be publicly accessible. The research demonstrates significant improvements in both qualitative and quantitative outcomes, presenting clinically plausible 3D reconstructions from ultrasound scans in diverse environments.

This project is led by Rishit Dagli, alongside Atsuhiro Hibi, Rahul G. Krishnan, and Dr. Pascal Tyrrell

Synovial Hypertrophy of the Knee – Generative AI

The purpose of the project was to help train deep learning models used for medical imaging. These models require extensive training data, which is difficult to obtain in medical imaging due to the high costs of data collection and labeling, the numerous restrictions on the distribution of patient data, and the rarity of certain diseases. To address this challenge, various methodologies have been proposed to enrich medical imaging datasets.

While augmentation is the traditional approach, modern methods increasingly rely on machine learning. This project aimed to compare traditional augmentation with a novel approach involving denoising diffusion probabilistic models (DDPMs). A DDPM was built and trained to generate new musculoskeletal ultrasound (MSK-US) images depicting thickened synovium and joint recess distension.

The performance of a convolutional neural network (CNN) was then compared when trained on unenriched data versus data enriched through either traditional augmentation or DDPM-generated images. The results demonstrated that the CNN’s diagnostic accuracy improved significantly when trained on the dataset enriched with DDPM-generated images, suggesting that this custom methodology is preferable over traditional augmentation.

Real Image

Synthetic Image (DDPM Generated)

Augmented image

This project is led by Benedek Balla, alongside Dr. Pascal Tyrrell, and Atsuhiro Hibi

PaddlesApp Stroke Count Algorithm 

In the Stroke/Peak Detection Algorithm project, the goal is to identify the most effective algorithm for detecting strokes and peaks using gyroscope and acceleration data. This study aims to validate the accuracy of the algorithms, which currently employ Fourier and Gaussian filtering techniques, in detecting stroke cycles. Accurate stroke detection is essential for ensuring that measurements reflect true changes in movement rather than inconsistencies in data collection. 

Data Collection: Participants are required to collect data by wearing the watch and following a strict protocol, which is in the process of being finalized. 

Algorithm Development: Active brainstorming is underway to improve stroke/peak detection and visualization. The current data lacks the precision needed for advanced visualizations in the application UI. 

Validation Plan: Access to a beep test, which provides expected stroke timing, would be invaluable for validation efforts. 

The focus will remain on accurately detecting the stroke cycle with simpler, more efficient algorithms. 

This project is led by Ramiz Khan, alongside Marshal Guo, Dr. Pascal Tyrrell, Minh Dang, and Ali Germany

Evaluating the Impact of Sex Bias on AI Models in Musculoskeletal Ultrasound Diagnostics for Synovial Recess Distension

The study examines how sex bias in AI training data affects the performance of AI models in diagnosing synovial recess distension using musculoskeletal ultrasound (MSKUS). Using a dataset of 5000 MSKUS images categorized by sex and clinical findings, two binary convolutional neural network (BCNN) classifiers were developed. The dataset was balanced across genders and conditions, with models trained using data augmentation and validated through a 5-fold cross-validation strategy.

AI models trained on female-only data showed higher sensitivity and accuracy but lower specificity for male images, indicating a bias. Models trained on balanced datasets demonstrated better generalizability. Classification heatmaps showed variations in model focus areas based on the training data’s demographic characteristics. The study highlights the critical impact of demographic biases on AI model performance in medical imaging, advocating for diverse and unbiased training datasets to improve diagnostic accuracy and reliability.

This work is the result of a collaborative effort between Shawn Demello, Dr. Pascal Tyrrell, Mauro Mendez, Christine Lee, and Minh Dang

Comparative Analysis of Adversarial Perturbation, Gaussian Noise, and Resolution Loss on Model Performance and Decision Making 

This study aims to examine and interpret the differential effects of adversarial perturbations, Gaussian noise, and resolution degradation on the performance and decision-making of a binary CNN classifier used for medical image classification tasks. The research involves systematically introducing each type of perturbation to the input images and plotting the resulting metrics against the level of perturbation. Additionally, the study explores how these perturbations affect the model’s decision-making process by utilizing Grad-CAM to visualize the areas of the images that the model focuses on. Through this analysis, the study seeks to gain insights into the robustness of the model and the nature of its decision-making under different types of perturbations. 

This project is led by Yuxi Zhu and Dr. Pascal Tyrrell, with special thanks to Daniel Saragih and Atsuhiro Hibi

Development and Evaluation of AI-Enhanced Point-of-Care Transcranial Doppler Ultrasound for Rapid Large Vessel Occlusion Detection in Acute Ischemic Stroke

An AI-enhanced ultrasound tool can significantly improve acute stroke care by enabling faster and more accurate identification of large vessel occlusions (LVOs), leading to more timely and appropriate treatment decisions. By optimizing the triage process and reducing time to treatment, this technology can enhance outcomes for stroke patients, particularly in areas where access to specialized stroke care may be limited.

This study aims to develop and evaluate an innovative AI-enhanced point-of-care transcranial Doppler ultrasound (POC TCD US) tool for rapid and accurate identification of LVO in acute ischemic stroke patients. The integration of machine learning algorithms with POC TCD US aims to standardize interpretation, enhance diagnostic accuracy and minimize operator dependence. By streamlining acute stroke triage and facilitating faster treatment initiation, this technology may transform acute stroke care. This method will optimize the triage process, ensuring patients requiring endovascular thrombectomy are promptly directed to comprehensive stroke centers while those suitable for intravenous thrombolysis are efficiently managed at primary stroke centers. This study will provide critical evidence on the effectiveness of AI-powered TCD in stroke management and contribute to the ongoing research and development of innovative diagnostic tools in this field.

This project is led by Eric Dhillon, Stephen Dillon, and Dr. Pascal Tyrrell

AI assessment of Ultrasound Adnexal Ovarian Masses

An ovarian adnexal mass is a common gynecologic issue which can either be benign or malignant, which can usually be distinguished by ultrasound features. Expert subjective assessment is considered the gold standard for diagnosis of benign or malignant lesion but is limited by the number of examiners. There is therefore some interest in automating this process with the use of machine learning and deep learning techniques to present expert readers with high priority cases that match algorithmic criteria suggesting malignancy, prompting faster review. We aim to develop an algorithm that classifies adnexal masses into benign or malignant categories using convolutional neural networks (CNN) to identify the highest priority cases that are most likely to represent malignancy, allowing for faster intervention of a deadly cancer.

Students on the project: Emma Blanchette, and Marshal Guo 

Evaluating the Effectiveness of LPIPS Versus SSIM and PSNR for Image Similarity Assessment Under Degradation Conditions

Image quality assessment (IQA) is a critical area of study, especially in the field of
medical imaging, where the accuracy and clarity of images are paramount for diagnosis and treatment planning. With the advent of advanced image-generating models, including diffusion models and Generative Adversarial Networks (GANs), there is a growing need to evaluate the effectiveness of the images generated by these models in maintaining or enhancing image quality. Traditional metrics like the Structural Similarity Index (SSIM),
Peak Signal-to-Noise Ratio (PSNR), and the more recent Learned Perceptual Image Patch Similarity (LPIPS) are commonly used to quantify image quality. This research uses these metrics to assess the impact of various degradation and determine which metric best
assesses similarities.

This project is led by Jingwen (Lisa) Zhong, and Dr. Pascal Tyrrell

Improving Performance Of ResNet18 Binary Classifier For Knee Recess Distension Using StyleGAN2 ADA Generated Images

This research addresses the challenges of using Convolutional Neural Networks (CNNs) for medical image classification, particularly in pediatric knee ultrasound images. While CNNs are commonly successful in image classification, they require large amounts of high-quality labeled data, which is difficult to obtain in medical imaging. Existing CNN-based models for knee recess distension, typically trained on adult data, underperform in children due to developmental differences in joint anatomy. Pediatric data, especially for hemophilia patients, is scarce, further complicating accurate classification.

To overcome this, the study explored using synthetic images generated by StyleGAN2 ADA to augment the training data for a ResNet18 model. These synthetic images were tailored to two pediatric age groups (0-8 and 9-13 years) and were added to the model’s training set. The results showed that the inclusion of these synthetic images significantly improved the model’s specificity, enhancing its performance in pediatric cases. This research demonstrates that GAN-generated data can effectively augment CNN training, improving image classification in medical contexts where real pediatric data is limited.

Authors: Karan Chahal, and Dr. Pascal Tyrrell
Affiliations: University of Toronto, SofTx Innovations

Investigating If False-Positives By An AI Breast Cancer Classification Model Are Associated With A Greater Risk For Breast Cancer Diagnosis In A Subsequent Screening

False-positives in any diagnostic assessment are undesirable, bringing anxiety and unnecessary follow-up assessments. But, in what cases might they hold useful information? Past studies have found that false-positives in mammogram readings by radiologists are associated with a greater risk for breast cancer development in the future. As the development of AI breast cancer classification models prosper in the field of mammography, this project aimed to investigate if similar findings may exist for false positives by AI breast cancer classification models.

Using a longitudinal mammography dataset, this study investigates the hazard ratio of receiving a breast cancer diagnosis in a screening subsequent to an exam of the same individual that received a false-positive result from an AI breast cancer classification model. This research is crucial, as such an association would suggest the possibility of using false-positives to stratify mammography screenings for individuals who are at higher risk of breast cancer development.

This project is led by Yan Qing Lee, and Dr. Pascal Tyrrell

Multi-Modal Learning for Ultrasound Imaging

Multi-modal learning leverages multiple information gained from different modalities for enhancing a better performance of detecting targets which a single modality could not have achieved. This concept, originating from machine learning, has been widely introduced in medical imaging, especially ultrasound imaging. In this project, we aim to review papers adopting multi-modal learning for ultrasound images and to identify useful algorithms or methods applicable to our future research.

Deep Generative Machine Learning

The purpose is to create a production pipeline for the artificial generation of ultrasound datasets in MSK. The proposed pipeline will generate synthetic data with the possibility of producing different stages of development of the disorder. Similarity measurement and comparison methodologies between authentic and synthetic data will be explored. The full title of this paper is “Deep Generative Machine Learning to Obtain Synthetic Ultrasound Data  for Rare Disease Computer Vision Research”.

Detection of Heterogeneity of CNNs

Convolutional Neural Networks (CNN) are highly dependent on an abundance of good-quality, labelled data for complete training. With small or overly heterogenous training sets, a CNN may report passable overall test accuracy, but perform poorly or inequitably for certain subgroups of data. Post-hoc cluster analysis of the feature space was shown to detect groups for which a CNN performs suboptimally, and that variation of CNN prediction is improved with greater data availability. Current work focuses on determining empirical benchmarks, model selection, and comparison of complementary methods (canonical correlation analysis). The full title of this project is “Cluster Analysis on the Detection of Heterogeneity of Convolutional Neural Networks”.

CAD/AI in Medical Imaging

Computer-aided diagnosis (CAD) is a technology that is rapidly becoming one of the most researched topics in radiology. The purpose of this systematic review is to describe and determine the facilitators and barriers to the introduction and adoption of Computer-aided diagnosis (CAD), a type of Artificial Intelligence (AI), by Radiologists in their workplace. We use literature covering the period from 2000 to 2020, and specific search strategies. The full title of this project is “Computer-Aided Diagnosis (CAD) and Artificial Intelligence (AI) in Clinical Practice: Systematic Review on the Barriers and Facilitators to CAD Integration in Radiology”.

Unknown Class Detection

Conventional deep rural network(DNN)-based methods take it for granted that prediction classes are the same between training, evaluation and deployment phases. The purpose of this research is to develop a method to incorporate into conventional DNN models that will allow for the identification of possible unknown classes (also termed, out of distribution), applicable to any medical imaging modalities, such as brain CT/MRI image datasets.

Automatic Detection & Segmentation

The full title of this project is “Automatic detection and segmentation of hemarthrosis on hemophiliac patients using ultrasound”. The project consists of designing, testing and deploying deep learning methodologies for the detection and segmentation of hemarthrosis in the joint space (knee, elbow and ankle) of hemophiliac patients. Funded by Novo Nordisk A/S.

16 Bit

16 Bit is a Toronto-based company founded by two radiology physicians with backgrounds in computer science and engineering. Using various world-class medical datasets, they design custom neural network architectures. With a unique combination of medical and technical expertise, 16 Bit focuses on solving the most impactful clinical problems facing medicine today.

Can-Tico

The aim of this initiative is to (1) establish a research agreement and network between the Costa Rican healthcare system and Canadian medical imaging research (MiDATA – UofT) to share and transfer data and knowledge; and (2) create a research environment and develop the methodological pipeline and associated tools required for transformative AI solutions for clinical applications involving medical imaging.

MiNE

The Medical image Network Enterprise (MiNE) is a scalable, community-defined e-infrastructure with the primary objective of housing an electronic image-based inventory and data warehouse. MiNE also provides related tools for research and education, to support and encourage the University of Toronto research community. MiNE advocates the efficient use and privacy-sensitive sharing of existing research and clinical image data.

Novo Nordisk

Novo Nordisk discovers and develops innovative biological medicines and makes them accessible to patients throughout the world. Their purpose is to drive change to defeat diabetes and other serious chronic diseases such as obesity, and rare blood and rare endocrine diseases by engineering scientific breakthroughs, expanding access to our medicines and working to prevent and ultimately cure the diseases they treat.

SofTx Innovations

SofTx Innovations is a software development startup company who builds deep tech applications, products, and platforms to integrate and leverage AI/ML advances for healthcare. As AI is a disruptive technology that provides unique insights, they help organizations identify where opportunities for continuous improvement exist in the areas of quality, productivity, and cost reduction across the organization.

IBM Watson Health

IBM works to enable healthcare technology to help organizations gain more insights to their data and simplify their operations. IBM Watson Health bridges data, technology and expertise together to transform healthcare. This allows for efficient, resilient and robust institutions . They provide solutions for healthcare providers, government, life sciences, diagnostic imaging, healthcare payers, and employers.

Nippon Steel

Nippon Steel looks to pursue world-leading technologies and manufacturing capabilities, and contribute to society with their products and services. They create and design various products like plates, flat products, bar and wire rod, construction products, pipes and tubes, stainless steel, titanium, railway, automotive and machinery parts.