Search

Browse

• Communities
• Authors
• Year
• Titles
• Subjects

My Space

• Login

Analysis and segmentation of coronary arteries using Novel Deep Learning techniques

Papamatthaiaki Ilektra-Despoina

Full record

---

Summary

Coronary Artery Disease (CAD) is among the principal causes of deathglobally. Since current diagnostic methods are primarily invasive, thereis an increasing interest in accurate, non-invasive alternatives. In thisregard, deep learning developments have proven pivotal in estimating di-agnostic indices. 3D modeling of coronary arteries in an accurate manneris crucial for enabling dependable, non-invasive diagnostic processes. Thisthesis explores the use of deep learning for coronary arteries segmentationin computed tomography angiography (CTA) images, with a view to-wards facilitating early diagnosis and improving treatments. Two modelsare compared and evaluated: Basic U-Net, a convolutional neural network(CNN), and a transformer-based model, UNETR. Both are implementedwithin the same framework so that a direct and fair comparison is ensured.Considering the increasing interest in transformer architectures withinthe medical field, this comparison intends to assess whether they holdtangible potential for improving the research in medical image segmen-tation. Although experimental results indicated that both models hadhigh accuracy, Basic U-Net performed consistently better than UNETR,especially when there were constraints of limited data and computationalresources. A quantitative evaluation using the Dice Similarity Coefficient(DSC) revealed an average score for Basic U-Net of 90.14%, in compar-ison to 89.56% for UNETR. Although theoretically, UNETR has an ad-vantage in capturing dependencies over greater distances, its performancewas likely constrained by its higher data requirements and sensitivity tocomputational limitations.These findings indicate that convolutional architectures remain morereliable under low-resource conditions and highlight the importance ofmodel selection and dataset size when it comes to medical applications inthis field. Overall, the research affirms that deep learning is feasible forcoronary arteries segmentation and further supports its potential applica-tion for non-invasive diagnostic processes for CAD.

Available Files

Services