Browsing by Author "LHR TP 8345"

Now showing 1 - 1 of 1

Automated White Matter Hyper Intensities Segmentation in MR Images
(Library Information Services, COMSATS University Islamabad, Lahore Campus, 2022) YAQEEN ALI; FA17-RCS-008; LHR TP 8345; Dr.Muhammad Aksam Iftikhar
White matter hyperintensities (WMHs) are high-intensity pixels in MRI, especially in FLAIR images. The presence of white matter hyperintensitiesn brain MRI is associated with cerebral small vessel diseases (CSVD), aging-related brain diseases, strokes, and dementia. The extraction of hyperrintensities from FLAIR MRI images of the brain is called WMH segmentation, which is important because it provides details about the volume, location, and shape of WMH lesions. These measurements and the quantity of WMHs are important in the research and diagnosis of a patient. In usual practice, the field expert (radiologist) performs manual segmentation of medical imaging, which is time consuming, costly, and subjective. A few years ago, several different semi-automated and fully automated methods were proposed for different tasks such as WHM detection, stroke lesion segmentation, and WMH segmentation. Some of these methods use supervised learning algorithms with handcrafted features or more recently learned features (representations). On the other hand, some methods also use unsupervised learning for segmentation. These days, deep neural networks are quite popular and robust for learning problems and claim more than human performance in several problems. A convolutional neural network is a deep network structure and is very attractive and effective in computer vision domain problems, especially as a robust method for image classification and segmentation. Recently, many algorithms have been proposed for WMH segmentation, including CNN, but the automatic WMH segmentation task is still challenging due to coexisting with other abnormalities and spatial variability of white matter lesions. This study used an ensemble model of three variants of the U-NET based models. A U-NET network is a network-advanced version of CNN for bio-medical image segmentation tasks. x Method The proposed architecture is an ensemble of three U-net base models. One model is fully convolution neural network and other two are Residual U-Net models. We create ensemble model by giving weights to three trained models. These weights are tuned through experiments. U-NET has two parts. One part is called a contracting path that learns features of an input image and reduces spatial information, and the other part is an expansive path that gradually combines features with spatial information from previous contracting path layers to produce a segmented mask of the input image. The proposed architecture uses reverse convolution or transpose of convolution operators in the second path instead of the convolution and pooling operations that are used in CNN and uses a large number of feature channels in the up-sampling part. Due to this, the model can better carry context information with spatial information to deeper layers of the network as compared to CNN. The residual blocks helps to train deep network without facing vanishing and exploding gradients problems. These characteristics of the proposed model makes it better than CNN and a simple U-NET for the WMHs segmentation problem. Result To evaluate the proposed system, the MICCIA WMHs Segmentation Challenge 2017 (http://wmh.isi.uu.nl/) dataset is used. Our study used some state-of-the-art evaluation measures like Dice Similarity Coefficient (DSC), a modified Hausdorff distance (H95), the absolute percentage volume difference (AVD), recall, and F1 Score. The DSC, AVD, H95, recall, and F1 scores respectively 0.93, 2.21, 5.13, 94%, and 93% against the ground truths for FLAIR and T1 sequences in the WMHs Segmentation Challenge 2017 dataset. Conclusion It is concluded that the proposed system provides a close match to expert delineation across the state-of-the-art dataset used by WMHs Segmentation Challenge 2017 training, and the proposed system is more efficient and effective for automated WHMs segmentation. The proposed system performs better in the standard evaluation measures than the state-of-art technique