Document Type : Full Research Paper
Authors
1 Department of Biomedical Engineering, Hamedan University of Technology, Hamedan, Iran
2 Department of Biomedical Engineering, Hamedan University of Technology
Abstract
Stroke is one of the causes of death and the main cause of disability in developed countries. Normally, identification of stroke lesions is done by magnetic imaging, and its analysis requires the continuous presence of a doctor in the treatment center. Therefore, intelligent processing of medical images will be an effective approach for automatic diagnosis of brain lesions.
In this paper, a new integrated framework based on fuzzy inference system and deep neural network for automatic segmentation of brain lesions is introduced. In this regard, firstly, an improved U-net deep network (U-net) has been introduced for lesion detection and segmentation, which includes increasing the number of encoder and decoder layers along with changing the activation functions. Then, by using a fuzzy inference system based on if-then rules used by membership functions, the proposed approach of this study, which is based on the pre-processing of input images and the use of the unit network, has been introduced.
The results showed that the integration of the fuzzy inference system in the pre-processing with the improved deep network could increase the DICE coefficient up to 0.84. In addition, improving the contrast of the input images by the fuzzy system compared to the usual pre-processing methods such as histogram equalization showed a much better performance in the detection of lesions with small dimensions, which is due to the ability to control the amount of contrast increase in the fuzzy systems compared to the usual methods.
Keywords
Main Subjects
)