Method and system for segmenting interventional device in image

A method of generating patches in an image of an interventional device for performing a deep learning image analysis of the image, the method comprising:

• dividing the image of the interventional device into a plurality of data units, each data unit having a plurality of pixels or voxels;
• determining a selected pixel or voxel of the plurality of pixels or voxels in each data unit to be a key point corresponding to the data unit by applying an operation to determine a rank order of the plurality of pixels or voxels based on a feature, and determining the selected pixel or voxel to be the key point according to the rank order;
• determining spatial positions of the key points in the image;
• determining strengths of the key points in the image, and distributions of the determined strengths of the key points;
• sorting the key points in a ranked order from most salient to least salient based on the determined strengths of the key points;
• computing a number N of salient key points from the determined strengths of the key points, the determined distributions of the determined strengths of the key points, and the determined spatial positions of the key points;
• selecting the number N of the sorted key points in the ranked order;
• and
• generating patches in the image using the selected key points.

A method of adapting a segmentation model for performing segmentation of an image, the method comprising:

• receiving a plurality of training images for training the segmentation model;
• generating device segmentation maps indicating an interventional device in each training image using semi-automatic shape registration;
• constructing ground truth data for the plurality of training images using the device segmentation maps;
• automatically adapting at least one of (i) a threshold to an imaging protocol for the segmentation model, (ii) input data to the segmentation model, or (iii) a parameter of the segmentation model by training the segmentation model using the ground truth data;
• acquiring a new test image showing the region of interest of the subject;
• and
• performing segmentation of the new test image including an interventional device using the adapted segmentation model.