Edge detection of MRI images using artificial neural network
Introduction Many methods have been proposed for MRI tissue segmentation. It has been identified that MRI image of human tissue is homogeneous and the structure of each is tissue connected, but it is rather difficult to separate the adjacent tissue due to the small intensity changes or smooth bound...
| Main Authors: | , , |
|---|---|
| Format: | Article |
| Language: | English English |
| Published: |
Malaysian Medical Association
2017
|
| Subjects: | |
| Online Access: | http://irep.iium.edu.my/59506/ http://irep.iium.edu.my/59506/ http://irep.iium.edu.my/59506/2/MJM_v72-Supp-1-2017.pdf http://irep.iium.edu.my/59506/8/59506_Edge%20Detection%20of%20MRI%20Images%20using_abstract_complete.pdf |
| Summary: | Introduction
Many methods have been proposed for MRI tissue segmentation. It has been identified that MRI image of human tissue is homogeneous and the structure of each is tissue connected, but it is rather difficult to separate the adjacent tissue due to the small intensity changes or smooth boundaries observed. Current traditional edge detection performance could be further improved using artificial neural network (ANN) based edge segmentation.
Methods
In this study, various existing edge detectors techniques based on spatial domain were evaluated. The best algorithm was selected subjectively, i.e. Canny edge detector, and is used as the training data for the proposed ANN. For each pixel of grayscale image, we obtain three features, i.e. horizontal (dx), vertical (dy), and diagonal (dz). The feedforward neural network was configured to have 1 input, 1 hidden, and 1 output. To determine the final pixel edge value (0 or 1), an optimum threshold was utilized.
Results
To obtain the best parameters which produce optimum edge of MRI images, we varied the number of neurons in the hidden layer, as well as the threshold. We found that the optimum parameter could be achieved by setting the number of neurons in the hidden layer to be 180, and the threshold to be 0.1 for various MRI images tested. We also found that the proposed ANN based edge detection has faster computation by almost three times compared to traditional Canny edge detector.
Discussion
The proposed ANN based edge detection produce better image segmentation for MRI images compared to other traditional edge detection algorithms. Moreover, the computational cost is smaller by almost three times compared to Canny edge detector. It is believe that such findings would allow medical practitioners to better obtain information that could be extracted from MRI images. |
|---|