Institutional Repository
Technical University of Crete
Technical University of Crete
EN
|
EL
| URI: | http://purl.tuc.gr/dl/dias/97234BFD-DFEF-4503-BC69-06200F50D07A | ||
| Year | 2025 | ||
| Type of Item | Diploma Work | ||
| License |
|
||
| Bibliographic Citation | Stelina Naka, "Analysis of cardiac anatomy biomedical images with the use of Vision Τransformers", Diploma Work, School of Electrical and Computer Engineering, Technical University of Crete, Chania, Greece, 2025 https://doi.org/10.26233/heallink.tuc.102213 | ||
| Appears in Collections |
The heart is one of the most complex organs of human body with multiple substructures and the anatomy of the whole heart is a basic requirement for the developing of many clinical applications. To study spatially heart function Magnetic Resonance Imaging (MRI) and Computed Tomography (CT) are the most common ways. In this regard, the whole-heart segmentation is vital in medical imaging analysis, providing the potential for diagnosis and treatment options of the Cardiovascular Diseases (CVD). However, the automated segmentation can be challenging due to variation of the heart shape. In this thesis, an enhanced method based on the insights of the MIC-CAI Multi-Modality Whole Heart Segmentation (MM-WHS) challenge evaluations is proposed. The challenge provides a dataset of 20 MRI and 20 CT volumes and their manually segmented labels. For later-on model training, most of the automated segmentation tasks on medical images are based on Convolutional Neural Networks, a hybrid Vision Transformer (ViT) model is introduced in this thesis. The so-called ’ViTSegment’, the proposed model, is a Vision Transformer-based encoder for capturing long range dependencies and a convolutional decoder for accurate boundary detection. The proposed algorithm was trained and evaluated on the dataset from the challenge. Due to the low number of data, we further proceeded with data augmentation techniques to expand the dataset. On CT dataset it exhibited a better dice score of 92.65 ± 2.17% compared to the MRI dataset,(91.50 ± 1.72%). To boost the results of the evaluation, a comparative analysis was implemented between the ViTSegment, U-Net and UNETR models. The ViTSegment outperforms the other two models, with U-Net achieving a dice score of 82.67 ± 8.70% on the CT dataset and 81.30 ± 5.47% on the MRI, while UNETR scores 86.33 ± 0.74% for CT and 84.94 ± 6.25% for MRI, highlighting its robustness and efficiency. ViTSegment model shows essential potential that is paving the way for robust automated whole-heart segmentation in medical image analysis.
Copyright © DIAS 2013
