The impact of learning rate, network size, and training time on unsupervised deep learning for intravoxel incoherent motion (IVIM) model fitting
Misha Pieter Thijs Kaandorp1,2, Frank Zijlstra1,2, João P. de Almeida Martins1,2, Christian Federau3,4, and Peter T. While1,2
1Department of Radiology and Nuclear Medicine, St. Olav’s University Hospital, Trondheim, Norway, 2Department of Circulation and Medical Imaging, NTNU – Norwegian University of Science and Technology, Trondheim, Norway, 3Institute for Biomedical Engineering, University and ETH Zürich, Zurich, Switzerland, 4AI Medical, Zürich, Switzerland
We demonstrate that a high learning rate, small network size, and early stopping in unsupervised deep learning for IVIM model fitting can result in sub-optimal solutions and correlated parameters. In simulations, we show that prolonging training beyond early stopping resolves these correlations and reduces parameter error, providing an alternative to exhaustive hyperparameter optimization. However, extensive training results in increased noise sensitivity, tending towards the behavior of least squares fitting. In in-vivo data from glioma patients, fitting residuals were almost identical between approaches, whereas pseudo-diffusion maps varied considerably, demonstrating the difficulty of fitting D* in these regions.
This abstract and the presentation materials are available to members only;
a login is required.