Purpose
Whole heart segmentation is an important method in medical imaging, used for modeling and analysis of heart anatomy and functions. Our aim is to assess the reliability of a semi-interactive method for whole-heart segmentation in MRI images of patients with atrial fibrillation (AF), as well as to improve the performance of current available solutions.
We propose a semi-interactive image segmentation algorithm based on region growing. Specifically, we extend a popular algorithm – GrowCut (Vezhnevets, Vezhnevets, & l., 2005) by endowing it with novel neighborhood structures...
Methods and Materials
Data were captured by a Discovery MR 750w General Electric 3.0T. The images were analyzed by a radiologist who manually segmented all four heart chambers. The results were controlled independently by a second radiologist. This resulted in one manual segmentation for each slice of each acquisition, which we treated as the ground truth. Fig. 1 shows a graphical overview of our workflow.
The raw images were then algorithmically segmented using our GrowCut extension. In the original approach, standard neighborhoods (Moore and Von Neumann) were used;...
Results
The data used consisted of stacks of 2D MRI images of 10 healthy controls and 10 AF patients participating in our imATFIB study (NCT03584126). Each volume consisted of 14-21 slices of 256x256 pixels. Each pixel measures 1.4 x 1.4 mm, with an 8 mm spacing between slices.
In our evaluation, we measured all VISCERAL reported metrics (Taha & Hanbury, 2015). We performed the evaluation in two regimes: the slice with the maximum foreground area (image-based) and stacks of images (patient-based). For the stacks of images,...
Conclusion
We extended the GrowCut image segmentation algorithm by endowing it with a global view of the image. We validated the algorithm on a Whole Heart Segmentation task applied to cardiac MRI images from our imATFIB study. By employing our algorithm in the medical image segmentation workflow, we gained several advantages:
Speedon large batches of imagesvs. the manual approach.
High-quality segmented images, validated by radiologists.
Improved performance over original GrowCut algorithm.
Consequently, our algorithm could be used as a building block in a Computer Aided Diagnostic...
Personal Information
Author affiliations
Radu Mărginean1,2, Loredana Popa1, Mihaela Coman1, Simona Manole1, Anca Andreica1,2, Laura Dioşan1,2, Zoltán Bálint1,3
1IMOGEN Research Institute, County Clinical Emergency Hospital
2Faculty of Mathematics and Computer Sciences, Babeș-Bolyai University
3Faculty of Physics, Babeș-Bolyai University, Cluj-Napoca, Romania
Acknowledgement
The authors wish to thank to Silviu Ianc and Cristina Szabo for their technical assistance, Dr. Claudia Budurea, Dr. Alin Iliescu and Dr. Sorin Pop for patient recruitment and cardiologic validation.
The authors highly acknowledge financial support from the Competitiveness Operational Programme 2014-2020 POC-A1-A1.1.4-E-2015, financed under...
References
Taha, A. A., & Hanbury, A. (2015). Metrics for evaluating 3D medical image segmentation: analysis, selection, and tool. BMC Medical Imaging, 15, 29. https://doi.org/10.1186/s12880-015-0068-x
Vezhnevets, V., & Konouchine, V. (2005). GrowCut -- Interactive Multi-Label N-D Image Segmentation By Cellular Automata. Retrieved from http://citeseerx.ist.psu.edu/viewdoc/summary?doi=10.1.1.59.8092