Abstract #1719

Comparison of radiomics-based and deep learning techniques for predicting nuclei density in brain cancer patients from MRI

Samuel Bobholz¹, Allison Lowman², Alexander Barrington³, Michael Brehler², Jennifer Connelly⁴, Elizabeth Cochran⁵, Anjishnu Banerjee⁶, and Peter LaViolette^2,3

¹Biophysics, Medical College of Wisconsin, Wauwatosa, WI, United States, ²Radiology, Medical College of Wisconsin, Wauwatosa, WI, United States, ³Biomedical Engineering, Medical College of Wisconsin, Wauwatosa, WI, United States, ⁴Neurology, Medical College of Wisconsin, Wauwatosa, WI, United States, ⁵Pathology, Medical College of Wisconsin, Wauwatosa, WI, United States, ⁶Biostatistics, Medical College of Wisconsin, Wauwatosa, WI, United States

This study sought to compare localized predictions of cellular density via radiomics-based and neural network-based modeling, using co-registered autopsy tissue samples from 16 brain cancer patients as ground truth. We found that radiomics models tended to slightly outperform the neural networks, despite evidence of overfitting in all radiomics-based models and an Alexnet-based transfer learning model. These results suggest that radiomics models tend to perform at least as well as neural network when applied to this dataset, but the propensity of these models for overfitting highlights further needs to be addressed with modelling on larger data sets.

This abstract and the presentation materials are available to members only; a login is required.

Join Here