References
Tang, M., Valipour, S., Zhang, V., Cobzas, D. and Jagersand, M.
A deep level set method for image segmentation,
DLMIA@MICCAI2014
Kiros R.,Pupuri K., Cobzas D., Jagersand M.
Stacked Multiscale Feature Learning for Domain Independent Medical Image Segmentation
MLMI@MICCAI2014
|
Description
We proposed (MICCAI17) a novel image segmentation approach
that integrates fully convolutional networks (FCNs) with a level set
model. Compared with a FCN, the integrated method can incorporate
smoothing and prior information to achieve an accurate segmentation.
Furthermore, different than using the level set model as a post-processing
tool, we integrate it into the training phase to fiee-tune the FCN. This
allows the use of unlabeled data during training in a semi-supervised
setting.
Fig 1: Overview of the proposed FCN-levelset model. The pre-trained FCN is re ned
by further training with both labeled (top) and unlabeled data (bottom). The level
set gets initialized with the probability map produced by the pre-trained FCN and
provides a refined contour for fine-tuning the FCN.
We also proposed (MLMI@MICCAI14) a framework for learning features from data itself at multiple scales and depth. Our method uses two layers of stacked dictionaries.
Our features can be easily integrated into classifiers or energy-based segmentation
algorithms. We test the performance of our proposed method
on two MICCAI grand challenges, obtaining the top score on VESSEL12 and competitive performance on BRATS2012.
This work is the winner of the VESSEL12 challenge.
Fig 2: Visualization of our feature learning approach. Each volume slice is scaled using
a Gaussian pyramid. Patches are extracted at each scale to learn a dictionary D using
OMP. Convolution is performed over all scales with the dictionary filters, resulting in k feature maps. After training the first layer, the feature maps can then be used as
input to a second layer.
|