Düzlemsel X-ışını görüntülerinden 3B hacim rekonstrüksiyonu için derin öğrenme teknikleri

X-ray is most often the preferred form of medical imaging due to its low cost. However, being 2D, it is not as effective as 3D imaging techniques when it comes to diagnosis. On the other hand, 3D imaging is expensive and hence not as ubiquitous as X-ray, with computed tomography posing higher radiation risks linked to tissue damage. Having said that, 3D reconstruction (3Dr) from X-ray seems like a plausible compromise. Conventional 3Dr methods, which rely on statistical templates for 2D/3D registration, often fail to account for patient-specific anatomical variations, leading to inaccuracies. This thesis introduces two novel deep learning techniques for 3Dr from a single X-ray, namely, X-ray to Volume (X2V) and X-ray to Bone (X2B). X2V, incorporating a vision transformer based encoder, reconstructs highly accurate 3D organ models, which capture individual anatomical variability. X2B applies this methodology to skeletal structures, addressing challenges such as overlapping Hounsfield Unit values in X-rays, which hinder the extraction of fine structural details. Furthermore, as extensions to X2B, X-ray to Bone Registration (X2BR) and Biplanar X-ray to Bone (X2B2) techniques integrate object detection, neural implicit modeling, and non-rigid registration to generate high-resolution 3Dr of ribs and vertebrae. We show that all four techniques are superior than the state of the art in the literature through chamfer distance, intersection over union, f-score, normal consistency metrics. By bridging the gap between traditional statistical methods and modern deep learning, this thesis enables more personalized and precise diagnostic and therapeutic solutions, with the potential of improving clinical workflows and patient outcomes.