Predicting limb loss in patients with diabetic foot disease on magnetic resonance imaging using deep learning models
File(s)
Author(s)
Ahmad, Manal
Type
Thesis
Abstract
Introduction
Diabetic foot disease (DFD) is a complex disease and is associated with lower limb amputation.
Magnetic resonance imaging (MRI) is commonly used in patients with DFD. Early detection of
surrogate markers on MRI may help to predict the risk of limb loss.
Aim
To explore the use of deep learning models to identify potential predictors on MRI for limb
loss in patients with DFD.
Methods
Four methodological approaches were undertaken:
1. 419 MRI reports were retrospectively reviewed to identify salient imaging
features, including collections, ulceration, osteomyelitis, diabetic
myopathy/sarcopenia, and Charcot foot. Multiple deep learning models were
applied to predict amputation outcomes.
2. Two-dimensional MRI images of the foot at the base of the first metatarsal were
classified as mild, moderate, or severe sarcopenia. 824 three-dimensional MRI
scans images were manually annotated, with inter-rater reliability assessed on a
subset of 50 images. Following preprocessing and augmentation, 1740 images
were used to train Vision Transformer and ResNet architectures.
3. MRI scans were manually segmented and modelled using nnU-Net, with a
pseudo fat fraction calculated from muscle and plantar fat signal intensities.
Results
XGBoost was the only statistically significant report-based model, achieving 63.4% accuracy
(p<0.001). Vision Transformer and ResNet-50/101 achieved accuracies of up to 80.7% for
sarcopenia classification, with good inter-rater agreement (κ=0.8). nnU-Net segmentation
had a mean DICE score 0.4. Sarcopenia severity was not a significant predictor of major
amputation or death; however, mild sarcopenia was associated with the composite outcome
of all amputations and death (OR 1.368, p=0.0377).
Conclusion
These findings suggest that while sarcopenia severity alone is not a robust surrogate marker
for limb loss in DFD, MRI-derived characteristics combined with clinical variables may improve
risk stratification.
Diabetic foot disease (DFD) is a complex disease and is associated with lower limb amputation.
Magnetic resonance imaging (MRI) is commonly used in patients with DFD. Early detection of
surrogate markers on MRI may help to predict the risk of limb loss.
Aim
To explore the use of deep learning models to identify potential predictors on MRI for limb
loss in patients with DFD.
Methods
Four methodological approaches were undertaken:
1. 419 MRI reports were retrospectively reviewed to identify salient imaging
features, including collections, ulceration, osteomyelitis, diabetic
myopathy/sarcopenia, and Charcot foot. Multiple deep learning models were
applied to predict amputation outcomes.
2. Two-dimensional MRI images of the foot at the base of the first metatarsal were
classified as mild, moderate, or severe sarcopenia. 824 three-dimensional MRI
scans images were manually annotated, with inter-rater reliability assessed on a
subset of 50 images. Following preprocessing and augmentation, 1740 images
were used to train Vision Transformer and ResNet architectures.
3. MRI scans were manually segmented and modelled using nnU-Net, with a
pseudo fat fraction calculated from muscle and plantar fat signal intensities.
Results
XGBoost was the only statistically significant report-based model, achieving 63.4% accuracy
(p<0.001). Vision Transformer and ResNet-50/101 achieved accuracies of up to 80.7% for
sarcopenia classification, with good inter-rater agreement (κ=0.8). nnU-Net segmentation
had a mean DICE score 0.4. Sarcopenia severity was not a significant predictor of major
amputation or death; however, mild sarcopenia was associated with the composite outcome
of all amputations and death (OR 1.368, p=0.0377).
Conclusion
These findings suggest that while sarcopenia severity alone is not a robust surrogate marker
for limb loss in DFD, MRI-derived characteristics combined with clinical variables may improve
risk stratification.
Version
Open Access
Date Issued
2026-01-11
Date Awarded
2026-05-01
Copyright Statement
Attribution-NonCommercial 4.0 International Licence (CC BY-NC)
License URL
Advisor
Davies, Alun
Rockall, Andrea
Shalhoub, Joseph
Sponsor
Imperial College Healthcare NHS Trust
Publisher Department
Department of Surgery & Cancer
Publisher Institution
Imperial College London
Qualification Level
Doctoral
Qualification Name
Doctor of Philosophy (PhD)