FBE- Biyomedikal Mühendisliği Lisansüstü Programı - Yüksek Lisans
Bu koleksiyon için kalıcı URI
Gözat
Sustainable Development Goal "Goal 9: Industry, Innovation and Infrastructure" ile FBE- Biyomedikal Mühendisliği Lisansüstü Programı - Yüksek Lisans'a göz atma
Sayfa başına sonuç
Sıralama Seçenekleri
-
ÖgeFully supervised and semi-supervised semantic segmentation of cardiac MR using deep learning( 2021-09-02) Bolhassani, Mahyar ; Öksüz, İlkay ; 504191413 ; Biomedical Engineering ; Biyomedikal MühendisliğiHeart diseases are one of the primary causes of death worldwide. A key factor to accurately and effectively treating heart diseases is to have quantified measures like high-quality images of the organ. When we provided physicians with medical scans, they can pinpoint the kind of abnormality in the heart. Cardiac Ultrasound, CT, and MRI scans are some of the modalities that we can leverage, while each modality has both advantages and disadvantages. Depending on the situation and the patients' condition, we can choose a preferred modality. We concentrate on cardiac MRI, which is a non-ionizing modality that constructs high-quality images. Segmentation of different heart areas in CMR scans such as myocardium mass, wall thickness, left ventricle (LV), right ventricle volume, and ejection fraction (EF) is a quantitative measure that assists cardiologists in diagnosing cardiac failures. Thanks to computer-aided detection (CAD) advancements, the automatic segmentation of the heart cavity for diagnosis purposes alleviates the burden of quantitative interpretation of large numbers of cardiac scans for cardiologists. The ultimate goal of training an automatic model is to predict correctly on unseen data. Therefore, we need a large number of labeled data which is a tedious and expensive task. However, the variation of CMR data acquisition from different centers or vendors demands us to have training data from almost all centers and vendors for a robust model, which is almost impossible. To address this issue, this thesis proposes a semi-supervised segmentation setup to leverage unlabeled data to segment the left ventricle, right ventricle, and myocardium regions. We utilize an enhanced version of residual U-Net architecture on a large-scale cardiac MRI dataset. Handling the class imbalanced data issue using dice loss, the improved supervised model can achieve better dice scores than a vanilla U-Net model. We applied standard augmentation techniques as well as histogram matching techniques to increase the performance of our model in the multi-domain setup. Also, we introduce a simple but efficient semi-supervised segmentation method to improve segmentation results without the need for extensive labeled data. Finally, we applied our approach to two benchmark datasets, STACOM LVQuan 18 and M\&Ms 2020 challenges, to show the potency of the proposed model. The quantitative results demonstrate the effectiveness of our proposed model. The model achieves average dice scores of 0.926, 0.933, and 0.892 for the left ventricle, right ventricle, and myocardium respectively.