Department

Mathematical Sciences

Author Type

Undergraduate Student

Submission Type

Event

Start Date

27-7-2017 1:15 PM

End Date

27-7-2017 2:15 PM

Description

The objective of the project is to develop practical algorithms to segment the human heart’s left and right ventricles from isotropic Magnetic Resonance Imaging (MRI). An advanced algorithm that allows for efficient analysis of the ventricles in 3D or 4D would help diagnose and treat numerous cardiac diseases and illnesses. The left ventricle is a convex-like structure whereas the right ventricle has a relatively more abstract shape. The constant blood flow in and out of the ventricles causes them to look different at any given time. Therefore, our data set is in 4D where the fourth dimension is time. Our algorithms take a 3D MRI volume and first reduces it to a 1D skeletal representation with figure-ground separation followed by morphological analysis. We think the skeletal representation is rich enough to keep the important structural information of ventricles while small enough to scale it to the 4D volume efficiently in the future. The first algorithm further reduces the skeleton into a minimum spanning tree and asks a user for an edge that can separate the ventricle from the rest of the structure. The second algorithm first asks the user for two points inside the ventricle: one near the atrium and the other near its tip. These two points define the skeleton of the ventricle and allows the algorithm to differentiate it from the other structures. We designed a user friendly graphical interface for our testing and demonstration. We tested the first algorithm on 4 data sets provided by the Fornwalt Lab at Geisinger Hospital. Each data contains manual annotations of both left and right ventricles at end-systolic and end-diastolic phases. The algorithm was run on a laptop with 1.80 GHz CPU with 8 GB RAM. The average Intersection over Union (IoU) and time was 60% and 1.484 seconds respectively for the left ventricle and 32% and 1.5281 seconds respectively for the right ventricle. We will report the results of the second algorithm at the conference. The target is 80% IoU for both ventricles.

Share

COinS
 
Jul 27th, 1:15 PM Jul 27th, 2:15 PM

Segmentation of Heart Ventricles Using Skeletal Representations

The objective of the project is to develop practical algorithms to segment the human heart’s left and right ventricles from isotropic Magnetic Resonance Imaging (MRI). An advanced algorithm that allows for efficient analysis of the ventricles in 3D or 4D would help diagnose and treat numerous cardiac diseases and illnesses. The left ventricle is a convex-like structure whereas the right ventricle has a relatively more abstract shape. The constant blood flow in and out of the ventricles causes them to look different at any given time. Therefore, our data set is in 4D where the fourth dimension is time. Our algorithms take a 3D MRI volume and first reduces it to a 1D skeletal representation with figure-ground separation followed by morphological analysis. We think the skeletal representation is rich enough to keep the important structural information of ventricles while small enough to scale it to the 4D volume efficiently in the future. The first algorithm further reduces the skeleton into a minimum spanning tree and asks a user for an edge that can separate the ventricle from the rest of the structure. The second algorithm first asks the user for two points inside the ventricle: one near the atrium and the other near its tip. These two points define the skeleton of the ventricle and allows the algorithm to differentiate it from the other structures. We designed a user friendly graphical interface for our testing and demonstration. We tested the first algorithm on 4 data sets provided by the Fornwalt Lab at Geisinger Hospital. Each data contains manual annotations of both left and right ventricles at end-systolic and end-diastolic phases. The algorithm was run on a laptop with 1.80 GHz CPU with 8 GB RAM. The average Intersection over Union (IoU) and time was 60% and 1.484 seconds respectively for the left ventricle and 32% and 1.5281 seconds respectively for the right ventricle. We will report the results of the second algorithm at the conference. The target is 80% IoU for both ventricles.