Healthy bone reconstruction through generative deep learning

Abstract

Artificial Intelligence (AI) is transforming the clinical practice of orthopedic surgeons by combining technology with their technical skills. AI can play a primary role in standardizing pre-surgical planning, assisting orthopedic surgeons in the decision-making process to minimize medical errors, guiding appropriate surgical management, and reducing the cost and duration of surgery through intelligent solutions in the field of orthopedics. The complexity and variability of the glenoid cavity anatomy have been a challenge for the medical community, especially in reconstructive surgical interventions such as shoulder arthroplasty. This surgery is recommended for the treatment of osteoarthritis (OA), a pathology defined by the progressive degeneration of the articular cartilage of the humeral head and glenoid, causing pain, stiffness, and limitation of movement. In preoperative planning, a 3D reconstruction of glenoid bone defects can play a fundamental role in the comprehension of the patient’s native anatomy and, consequently, assist the orthopedic surgeon in the decision-making process, to restore the morphological parameters of the scapula, which is crucial for functional outcomes and the longevity of the implant. The main objective of this dissertation is the reconstruction of the healthy anatomy of the glenoid from three-dimensional computed tomography (3D CT) images through Generative Deep learning (GDL). In quantitative terms, the goal of this project is to virtually reconstruct glenoid bone defects so that the estimated version is within the range [-5º, 10º] since the purpose of shoulder arthroplasty is to accurately restore a healthy patient’s anatomy. This project explores two approaches for training a 3D image-to-image translation model: Pix2Pix and CycleGAN. In Pix2Pix, a reference image in the original domain 𝑋 is available for each image in the target domain 𝑌, allowing one-to-one mapping. In contrast, CycleGAN performs training with unpaired data, the images in domain 𝑋 are semantically related to the images in domain 𝑌, and there is not necessarily a reference image in domain 𝑌 for each image in domain 𝑋. The distinguishing feature of CycleGAN is the incorporation of cycle consistency loss, which facilitates training without paired data. In other words, this model translates from the original domain to the target domain without a one-to-one mapping. This study aims to investigate and compare the performance of these two architectures in the context of healthy bone reconstruction. Concisely, the generative model (CycleGAN or Pix2Pix) seeks to learn the mapping function between two domains, 𝐺 ∶ 𝑋 → 𝑌, i.e., the conversion of an image 𝑥 from the domain 𝑋 to an image 𝐺(𝑥) from the domain 𝑌, where 𝑥 is an image of a scapula with the glenoid removed and 𝐺(𝑥) the sample produced by the generator, which wants to conceive images similar to those of the real dataset of the healthydomain. The study demonstrates the potential of the CycleGAN and Pix2Pix models to reconstruct a healthy bone from a defective bone. Taking into consideration a significantly larger dataset, both models are expected to outperform in reconstructing a defective glenoid. That opens up a possibility for the development of an automated and intelligent virtual reconstruction tool that can be used in clinical applications, to ensure that the preoperative planning process of shoulder arthroplasty is efficient and quick, guide an appropriate surgical management, facilitate communication between surgeons, minimize medical errors, provide prognostic information, and optimize the performance of shoulder arthroplasty.