Generation of synthetic segmented 3D-CT scan of healthy Pelvic bone

Deep neural networks have provided notable progress in the field of medical image analysis. However, owing to their insatiable demand for data and the relatively constrained datasets in medical imaging projects, could limit their full potential. Medical Data is often limited due to ethics and data protection regulations. Synthetic data emerges as a promising alternative, offering the opportunity to increase the diversity and quantity of training data to improve machine learning algorithms' performance and generalization capabilities. In this work, we used a set of 436 segmented Computed Tomography (CT) scans to train a latent diffusion model (LDM) and to learn the distributions of healthy pelvis images in order to create synthetic segmented CT images of pelvic regions for bone surgical planning purposes. This LDM model is based on a combination of an autoencoder and a diffusion model. Qualitative and quantitative analysis proves that our trained model can generate realistic segmented CT scans, retain morphological information and anatomical details of healthy subjects, and outperform recent approaches used as a baseline.