Automatic segmentation of the prostate is a crucial step in the computer-aided
diagnosis of prostate cancer and in treatment planning. Current methods for prostate
segmentation primarily rely on deep learning models with neural networks. However,
these models tend to be large and lack transparency, which is essential for
physicians. We proposed a new data-driven 3D prostate segmentation method on
MRI named Green U-shaped Learning (GUSL). Different from deep learning based
methods, GUSL employs a feed-forward system that utilizes successive subspace
learning (SSL).
To keep enough detailed information on a dataset with a large image size, we
propose a cascading model in two stages, as shown in Figure 1: (1) segmentation on
downsampled images and (2) segmentation on the cropped patches. GUSL consists
of three main modules, as shown in Figure 2: representation learning, feature
learning, and residual correction. All modules are applied at multiple levels with
varying resolutions. We achieve fine-to-coarse unsupervised representation learning
using cascaded VoxelHop units, as well as coarse-to-fine segmentation through
feature learning and residual correction. GUSL maintains a very competitive standing
performance-wise with other DL baseline models and keeps a smaller model size
and less complexity, with transparency for doctors.
