Dissertation Defense - Gökçe Güven (PHDCS)

Gökçe Güven – PhD, Computer Science 

Prof. Dr. Hasan Fehmi Ateş – Advisor

Date: 20.01.2025

Time: 15:00-17:00

Location: AB1- 407

“DEEP LEARNING TECHNIQUES FOR 3D VOLUME RECONSTRUCTION FROM PLANAR X-RAY IMAGES”

Prof. Dr. Hasan Fehmi Ateş

Prof. Dr. Hasan Fatih Uğurdağ, Özyeğin University

Prof. Dr. Tanju Erdem, Özyeğin University

Prof. Dr. Sezer Gören Uğurdağ, University of of Massachusetts Dartmouth

Prof. Bahadır Kürşat Güntürk, İstanbul Medipol University

Abstract:

This thesis introduces the X2V network, an innovative method for reconstructing three-dimensional (3D) organ volumes from single planar X-ray images. The proposed approach is particularly significant for clinical applications such as real-time image-guided radiotherapy, computer-aided surgery, and patient monitoring. Traditional 3D reconstruction techniques rely on statistical 3D organ templates for 2D/3D registration, which may not account for individual variations in organ shape. Our X2V model addresses this limitation by utilizing neural implicit representation and incorporating a vision transformer model as an encoder to enhance attention to specific regions within the X-ray image. The resulting reconstructed meshes closely resemble the actual organ volumes, demonstrating X2V's capability to accurately capture 3D structures from 2D images. Furthermore, the X2V network's methodology is extended to the reconstruction of 3D bone volumes from single X-ray images, introducing the X2B network. This network is designed to extract 3D structures of ribs, costal cartilages, the sternum, and vertebrae using a combination of object detection, neural implicit modeling, and non-rigid registration techniques. The X2B network addresses the challenges posed by overlapping Hounsfield Unit (HU) values in X-ray images, which complicate the extraction of high-resolution 3D details for complex structures like the spine. The enhanced X2B model utilizes non-rigid registration to accurately capture vertebral details and generate detailed 3D skeletal structures from digitally reconstructed radiographs (DRRs). The results demonstrate that the X2B network can effectively reconstruct detailed and accurate 3D bone volumes from single X-ray images, making significant advancements in medical imaging and diagnostics.

Bio:

Gökçe Güven is a PhD Candidate at Ozyegin University, specializing in advanced AI and computational techniques for biomedical applications. Gökçe's research background includes roles at TÜBİTAK, Özyeğin University, and the Sabanci University, where she worked on projecAB1-407 ts such as the structural analysis and classification of protein conformational motions in molecular systems. She holds a PhD in Computer Science from Özyeğin University and an MSc in Materials Science and Engineering from Sabancı University. Throughout her career, Gökçe has shared her findings at various international conferences and has several publications in high-impact journals.