Enhanced segmentation of femoral bone metastasis in CT scans of patients using synthetic data generation with 3D diffusion models

TL;DR

An automated data synthesis pipeline using 3D Denoising Diffusion Probabilistic Models (DDPM) is proposed to enchance the segmentation of femoral metastasis from CT-scan volumes of patients to show that segmentation models trained with synthetic data outperformed those trained on real volumes only, and that those models perform especially well when considering operator variability.

Abstract

Purpose: Bone metastasis have a major impact on the quality of life of patients and they are diverse in terms of size and location, making their segmentation complex. Manual segmentation is time-consuming, and expert segmentations are subject to operator variability, which makes obtaining accurate and reproducible segmentations of bone metastasis on CT-scans a challenging yet important task to achieve. Materials and Methods: Deep learning methods tackle segmentation tasks efficiently but require large datasets along with expert manual segmentations to generalize on new images. We propose an automated data synthesis pipeline using 3D Denoising Diffusion Probabilistic Models (DDPM) to enchance the segmentation of femoral metastasis from CT-scan volumes of patients. We used 29 existing lesions along with 26 healthy femurs to create new realistic synthetic metastatic images, and trained a DDPM to improve the diversity and realism of the simulated volumes. We also investigated the operator variability on manual segmentation. Results: We created 5675 new volumes, then trained 3D U-Net segmentation models on real and synthetic data to compare segmentation performance, and we evaluated the performance of the models depending on the amount of synthetic data used in training. Conclusion: Our results showed that segmentation models trained with synthetic data outperformed those trained on real volumes only, and that those models perform especially well when considering operator variability.

Figures (5)

• Figure 1: Examples of femoral CT scans from three patients with metastasis in coronal view with operator segmentations (first annotation in red and second annotation few months later in green for beginners)
• Figure 2: Proposed fully automated image synthesis pipeline of CT femoral metastasis used to generate realistic pathological images using existing lesions and healthy femurs.
• Figure 3: Example of data obtained with our synthesis pipeline. Real metastatic volume (A) with its associated segmentation (in red) is used in conjunction with healthy femurs (B) to create two distinct new pathological volumes (C). After using the Denoising Diffusion Probabilistic Model (DDPM), the volumes (D) are obtained.
• Figure 4: Examples of images obtained from two synthetic volumes with artificial lesion located in the greater trochanter on the top row and in the femoral neck on the bottom row after sampling for different values of timesteps $\lambda$.
• Figure 5: Examples of segmentations of two real CT-scans of femurs with metastasis using the five trained models. The ground truth is in red, the automatic segmentations are in green. One can note the large mistakes made by the networks on the lower end of the diaphysis on the three last columns.