Landmark detection with spatial and temporal constraints in medical imaging
Anatomy, such as papillary muscle, is automatically detected ( 34 ) and/or detected in real-time. For automatic detection ( 34 ) of small anatomy, machine-learnt classification with spatial ( 32 ) and temporal (e.g., Markov) ( 34 ) constraints is used. For real-time detection, sparse machine-learnt detection ( 34 ) interleaved with optical flow tracking ( 38 ) is used.
1. A method for detecting sub-valvular location in medical imaging, the method comprising:
acquiring a sequence of frames of medical imaging data, each of the fames representing a volume of at least a region of a heart of a patient;
detecting a valve bounding box enclosing a valve of the heart in each of the frames;
detecting, in each of the frames, a sub-valvular bounding box enclosing the sub-valvular structure location based on the valve bounding box;
detecting, in each of the frames, candidate locations of sub-valvular location within the sub-valvular bounding box;
removing the candidate locations outside a spatial constraint;
selecting one of the candidate locations remaining after the removing in each of the frames as the sub-valvular location, the selecting being a function of the candidate locations across the frames; and
highlighting the sub-valvular location in an image of at least the region of the heart of the patient.
2. The method of claim 1 wherein acquiring comprise acquiring transesophageal echocardiography data.
3. The method of claim 1 wherein detecting the valve bounding box comprises detecting a rectangular prism with a marginal space learnt classifier.
4. The method of claim 1 wherein detecting the sub-valvular bounding box comprises detecting a rectangular prism based on an average position of the sub-valvular bounding box with respect to the valve bounding box.
5. The method of claim 1 wherein detecting the valve bounding box comprises detecting a plurality of candidate mitral valve bounding boxes in each frame, and wherein detecting the sub-valvular bounding box comprises detecting a respective plurality of candidate papillary bounding boxes in each frame and selecting the papillary bounding box for each frame as a Random Sample Consensus of the candidate papillary bounding boxes across the frames.
6. The method of claim 1 wherein detecting the candidate locations comprises detecting with a marginal space learnt classifier.
7. The method of claim 1 wherein removing comprises limiting the candidate locations to a bounded region relative to a landmark of the valve the valve comprising a mitral valve.
8. The method of claim 1 wherein selecting comprises selecting with a Markov constraint.
9. The method of claim 1 wherein selecting comprises selecting a set from different sets of the candidate locations with only one of the candidate locations in each frame, the selecting being a function of temporal displacement between the candidate locations within the set.
10. The method of claim 1 wherein highlighting comprises displaying a graphic of a model of a mitral valve including a representation of a papillary muscle location and chordae connecting the papillary muscle location to leaflets of the mitral valve or displaying a graphic of the papillary tips without graphics for other valve structure.
11. The method of claim 1 wherein acquiring comprises selecting a sparse sampling of the sequence of a set of frames representing the heart over time;
further comprising locating the sub-valvular location in the frames not in the sparse sampling using optical flow.
12. The method of claim 11 further comprising periodically updating the sub-valvular location for the optical flow with the selected sub-valvular location.