Matt Hemsley1,2, Rachel W Chan2, Liam Lawrence1,2, Sten Myrehaug3,4, Arjun Sahgal2,3,4, and Angus Z Lau1,2
1Medical Biophysics, University of Toronto, Toronto, ON, Canada, 2Department of Physical Sciences, Sunnybrook Research Institute, Toronto, ON, Canada, 3Department of Radiation Oncology, Sunnybrook Health Sciences Centre, Toronto, ON, Canada, 4Department of Radiation Oncology, University of Toronto, Toronto, ON, Canada
1Medical Biophysics, University of Toronto, Toronto, ON, Canada, 2Department of Physical Sciences, Sunnybrook Research Institute, Toronto, ON, Canada, 3Department of Radiation Oncology, Sunnybrook Health Sciences Centre, Toronto, ON, Canada, 4Department of Radiation Oncology, University of Toronto, Toronto, ON, Canada
Neural network models were compared to a non-linear least
squares fitting of the Bloch-McConnell equations for reconstruction of quantitative magnetization
transfer images. Runtime was reduced from 10 hours/slice to 1 second/slice with
approximately 1% error.
Fig
2. Example
M0b outputs. The first column shows the output of the non-linear least-squares
fit of the Bloch McConnell equations used as the ground truth. Subsequent columns
show the ANN, CNN, and Hybrid model outputs. The second row contains the difference
maps of the associated model and Bloch-McConnell fit.
Fig 3. Model performance. A) Example segmentation of GTV and
cNAWM. B) and C) show Bland Altman plots of the single slice in A, and the full
test set, respectively. Scatterplots show the relationship between network
predictions and the non-linear least-squares fit of the Bloch-McConnell
equations, and the Pearson correlation coefficient for each ROI. The difference
plots evaluate the bias (solid line, dashed lines 95% CI) between mean differences
of the network and ground truth in each ROI.
