Ground truth
Model prediction (~150K parameters)
The automatic segmentation of medical images aims to associate a pixel to a label within a medical image without human interaction. Convolutional neural networks (CNNs) and deep learning have helped facilitate enormous strides in the efficacy of automatic segmentation for medical imaging over the last several years. However, modern deep learning paradigms for 3D medical image segmentation are computationally expensive, often requiring high-end computing resources and several days of training time on those resources. My research explores ways to reduce this computational cost while also preserving segmentation accuracy. By exploiting similarities between CNNs and numerical discretizations for solving partial differential equations, we can create neural networks capable of achieving comparable segmentation accuracy to conventional CNNs while requiring substantially fewer parameters. The figure above highlights the results of our model for the task of labeling a brain tumor (shown in red). A preprint of these results is available here.