Top-K Maximum Intensity Projection Priors for 3D Liver Vessel Segmentation

TL;DR

The paper tackles 3D liver-vessel segmentation by preserving global vessel topology through CT reconstruction physics. It introduces top-$k$ maximum intensity projections (top-$k$ MIP) as a conditioning signal for a latent diffusion model to recover the 3D vessel tree from integral projections $x_0$, with conditioning encoded as $c'$. Key contributions include a principled topology-encoding via top-$k$ MIP, a diffusion framework conditioned on $c'$, and an artifact-suppression step enforcing projection consistency, demonstrated on the 3D-ircadb-01b23 dataset. Results show competitive Dice, IoU, and sensitivity with improved vessel completeness, while limitations include memory constraints and challenges in extremely low-contrast regions. Overall, the work provides a physics-informed, diffusion-based approach to robust 3D liver-vessel segmentation with potential clinical impact for preoperative planning.

Abstract

Liver-vessel segmentation is an essential task in the pre-operative planning of liver resection. State-of-the-art 2D or 3D convolution-based methods focusing on liver vessel segmentation on 2D CT cross-sectional views, which do not take into account the global liver-vessel topology. To maintain this global vessel topology, we rely on the underlying physics used in the CT reconstruction process, and apply this to liver-vessel segmentation. Concretely, we introduce the concept of top-k maximum intensity projections, which mimics the CT reconstruction by replacing the integral along each projection direction, with keeping the top-k maxima along each projection direction. We use these top-k maximum projections to condition a diffusion model and generate 3D liver-vessel trees. We evaluate our 3D liver-vessel segmentation on the 3D-ircadb-01 dataset, and achieve the highest Dice coefficient, intersection-over-union (IoU), and Sensitivity scores compared to prior work.

Figures (6)

• Figure 1: (a) Standard CT reconstruction: Given the integral projections and projections directions $\overrightarrow{d}$, it reconstructs the underlying 3$D$ object by back-projecting the integral along each direction; (b) Our proposed top-k MIP: Given the projections directions $\overrightarrow{d}$, our model reconstructs the 3$D$ liver-vessel tree by computing the top-k maximum value of the CT scan along each direction and inputting this into a latent diffusion model.
• Figure 2: Model outline. Our model represents the global 3$D$ topology of the liver by computing the top-k MIP over the CT volume, $\mathbf{c}$. Subsequently, it encodes these top-k MIP into a latent condition, $\mathbf{c}^\prime$ (orange). This $\mathbf{c}^\prime$ latent is used to condition the latent diffusion model (blue) which recovers the ground truth vessel tree ${\mathbf{x}}_0$ from noisy inputs ${\mathbf{x}}_t$. We represent the ground truth via integral projections (IP) of the 3$D$ ground truth vessel tree. The ground truth ${\mathbf{x}}_0$ and the noisy input ${\mathbf{x}}_t$ are encoded via a KL-autoencoder (green) to be used in the latent diffusion U-net b19. We denote the different viewing directions by $V$ in the batch size.
• Figure 3: (a) Ground truth IP: We include all vessel tree slices in the integral projection (IP) and accumulate them along projection directions $\overrightarrow{d}$. (b) Condition top-k MIP: We include all CT slices in the top-k maximum intensity projection (top-k MIP) and keep the top-k maxima along projection directions $\overrightarrow{d}$;
• Figure 4: Qualitative comparison on the 3D-ircadb-01b23 dataset. We highlight in green the predictions of the 3$D$ methods and in yellow the predictions of the 2$D$ methods. Our model makes more complete and continuous predictions, resembling the ground truth.
• Figure 5: Vessel reconstruction with/ without artifacts suppression. (a) reconstructed vessel tree with artifact suppression; (b) reconstructed vessel tree without artifact suppression; (c) full-view projection of a slice of the reconstructed vessel trees with artifacts suppression; (d) full-view projection without artifacts suppression. Artifact suppression improves the consistency between different projection views. The vertical axis in (c) and (d) is the projection view.