Automated segmentation using deep learned priors
AssigneeGeneral Electric Company
InventorsSuvadip Mukherjee, Roshni Bhagalia, Xiaojie Huang
Life of this patent
Estimating the term…
Abstract
Embodiments described herein provide a hybrid technique which incorporates learned pulmonary nodule features in a model based energy minimization segmentation using graph cuts. Features are extracted from training samples using a convolutional neural network, and the segmentation cost function is augmented via the deep learned energy. The system and method improves segmentation performance and more robust initialization.
Claims19
A system comprising a hybrid method to process image analytics posed in a cost function minimization framework, the system comprising:
- one or more image data sets comprising image-based data that utilizes low level image features;
- a probability map that provides prior knowledge about an object;
- wherein the hybrid method comprises:
- providing at least one of the image-based data, providing a regularizing energy that imparts smoothness, and integrating the prior knowledge about the object via the probability map.
A non-transitory computer readable medium herein described comprises computer-readable instructions stored therein for causing a control processor to perform operations to determine a graph cut energy function comprising the steps of:
- determining a data term which is a function of a signed radial gradient of an image;
- using a deep learned probability map to segment the image to create a segmented image;
- and
- creating a graph by connecting positive voxels to a sink node, and connecting negative voxels to a source node;
- wherein the graph energy function utilizes a threshold surface in the segmented image, reducing error at the edges of the probability map, and providing enhancement of an object in the image.
A system comprising:
- an object segmentation module;
- one or more graph cuts and level sets, in combination, which are sensitive to seed localizations;
- and a deep learning module comprising a processor that learns discriminative features of an object and produces an object probability map that incorporates the object segmentation module.
Dependencies are inferred from the claim text.