Research
Machine learning for diabetic foot ulcers, generative models for medical images, population evidence for limb-preservation care, and quantitative MRI. Every paper has a plain-language summary.
Showing work tagged . Show all
-
2026
Temporal machine learning framework for diabetic foot ulcer healing trajectory prediction
Journal article. BioMedical Engineering OnLine, 25(1), 41 Open access
Wound care is mostly reactive: treatment changes after a wound has already stalled. This framework looks one appointment ahead. Using only the clinical notes clinics already record, with no imaging, it predicts whether a diabetic foot ulcer will move in a favourable, acceptable or unfavourable direction by the next visit, and pairs that forecast with treatment suggestions.
-
2025
Multimodal Artificial Intelligence System to Enhance the Classification and Analysis of Diabetic Foot Ulcer Healing Conditions
PhD thesis. University of Toronto Open access
The doctoral work that ties the research programme together. It builds the Zivot dataset, ranks the clinical signs that matter most for healing, measures how colour, thermal and depth imaging complement each other, and introduces an attention-based network that fuses all of them to classify a wound's healing phase.
-
2025
Applications of Generative AI in Diabetic Foot Ulcer Treatment
Book chapter. IntechOpen Open access
A single-author review of how generative AI is changing diabetic foot ulcer care: synthetic images where real data is scarce, automated clinical documentation, treatment planning and healing forecasts. It also covers what clinical validation and the new regulatory guidance for AI-enabled devices mean for getting these tools into practice.
-
2025
Accessible healing phase classification of diabetic foot ulcer
Journal article. Computers in Biology and Medicine, 197, 111066
Knowing whether a wound is in the inflammatory, proliferative or remodelling phase steers treatment, but judging it takes specialist experience. This study shows a machine learning model can make that call from routine patient and wound information, and uses explainable AI to show which clinical features drive the decision.
-
2025
Enhancing Diabetic Foot Ulcer Assessment Through Fine-Tuned Vision-Language Models
Conference paper. IJCAI 2025 workshops, Communications in Computer and Information Science (Springer), 117-131
General-purpose vision-language models can describe a photograph, but not the way a wound-care clinician would. This work fine-tunes vision-language models on diabetic foot ulcer images so they assess and describe wounds in clinical terms, moving from the off-the-shelf approach of UlcerGPT to models adapted to the domain.
-
2024
UlcerGPT: A Multimodal Approach Leveraging Large Language and Vision Models for Diabetic Foot Ulcer Image Transcription
Conference paper. ICPR 2024 workshops (PRHA), Lecture Notes in Computer Science (Springer), 221-232
UlcerGPT pairs a vision-language model (LLaVA) with a large language model (ChatGPT) to turn a photo of a diabetic foot ulcer into a written clinical description: what is in the image, where the regions of concern are, and how they should be classified. Expert clinicians rated its output on a public dataset.
-
2024
Effectiveness of a Multidisciplinary Limb Preservation Program in Reducing Regional Hospitalization Rates for Patients With Diabetes-Related Foot Complications
Journal article. The International Journal of Lower Extremity Wounds, 24(1), 117-123 Open access
An eleven-year look at Alberta's health records comparing regions served by the podiatric-surgeon-led Toe and Flow limb-preservation model with regions on standard care. Patients with diabetic foot complications in Toe and Flow regions were hospitalized less often and went home sooner.
-
2024
Protocol for metadata and image collection at diabetic foot ulcer clinics: enabling research in wound analytics and deep learning
Journal article. BioMedical Engineering OnLine, 23(1), 12 Open access
Good wound AI starts with good data. This paper sets out a standard way to collect diabetic foot ulcer data at the clinic, including a dedicated case report form, and introduces the resulting Zivot dataset: colour photos with matching thermal and depth maps and full clinical metadata. A baseline detector trained on it shows the data is clean enough to learn from.
-
2023
Synthesizing Diabetic Foot Ulcer Images with Diffusion Model
Conference paper. ECML PKDD 2023 workshops (ARIAL), Communications in Computer and Information Science (Springer), 353-360
Medical AI is starved of data, and wound images are especially scarce. A diffusion model trained on 2,000 diabetic foot ulcer photos produced synthetic wounds realistic enough that clinicians took them for real 70% of the time. The study also shows that the standard image-quality scores, FID and KID, do not agree with clinicians, so better ways to judge synthetic medical images are needed.
-
2022
Domain-Specific Deep Learning Feature Extractor for Diabetic Foot Ulcer Detection
Conference paper. 2022 IEEE International Conference on Data Mining Workshops (ICDMW), 1-5
Every wound detector is built on a feature extractor, the backbone that turns pixels into patterns. This study benchmarks 14 backbone and architecture combinations on the public DFU2020 dataset to find the best foundation for automated ulcer detection.
-
2021
Reduction in Diabetes-Related Major Amputation Rates After Implementation of a Multidisciplinary Model: An Evaluation in Alberta, Canada
Journal article. Journal of the American Podiatric Medical Association, 111(4)
Calgary and Edmonton share a provincial health system and similar demographics, but only Calgary runs the multidisciplinary Toe and Flow model. Ten years of amputation records show Calgary performing far fewer major amputations, with more limb-sparing minor procedures taking their place: evidence for rolling the model out across Canada.
-
2019
Transverse relaxometry with transmit field-constrained stimulated echo compensation
Journal article. Magnetic Resonance Materials in Physics, Biology and Medicine (MAGMA), 32(6), 669-677
Quantitative T2 maps can reveal subtle tissue changes, but real scanners never produce the clean exponential decay the standard fit assumes. Stimulated echo correction fixes the bias at the cost of noisier estimates. This paper's two-step method uses the fact that the transmit field varies smoothly across the head, fixing it in a second pass so each voxel fits fewer free parameters.
-
2019
Focal epilepsy without interictal spikes on scalp EEG: A common finding of uncertain significance
Journal article. Epilepsy Research, 150, 1-6
Doctors use interictal spikes on EEG to locate and classify epilepsy, but not every patient shows them. Reviewing seven years of video-EEG monitoring at the Calgary Epilepsy Program, this study measured how often adults with focal epilepsy never spike on scalp EEG and whether they differ clinically from those who do.
-
2018
Method for Optimizing Quantitative Temporal Lobe Epilepsy MR Imaging
MSc thesis. University of Calgary Open access
Standard MRI can miss the subtle hippocampal changes behind temporal lobe epilepsy. This thesis develops a leaner stimulated echo correction for T2 relaxometry and tests it on simulated and patient data, with the goal of making those lesions easier to find.
-
2016
Method for Optimizing Quantitative Temporal Lobe Epilepsy MR Imaging
Conference abstract. Calgary Epilepsy Research Retreat
An early report of the optimized T2 relaxometry method and its use for temporal lobe epilepsy imaging, presented to Calgary's epilepsy research community.
-
2016
Transverse relaxometry with B1+ constrained stimulated echo correction
Conference abstract. Proceedings of the 24th Annual Meeting of ISMRM, Singapore, 1529
The first presentation of the transmit-field-constrained fitting idea at the International Society for Magnetic Resonance in Medicine, showing that the variance of T2 estimates can drop while accuracy is kept.
Citation counts and the full record are on Google Scholar and ORCID.