Scale-specific auxiliary multi-task contrastive learning for deep liver vessel segmentation

TL;DR

The paper tackles automated liver vessel segmentation with an emphasis on preserving the multi-scale geometry of hepatic vasculature. It introduces a scale-aware framework that decomposes the vascular tree into scale levels via clustering and uses scale-specific auxiliary tasks together with a contrastive loss within a 3D UNet-like encoder-decoder to shape shared representations. The approach defines a multi-scale clustering scheme to generate scale masks and trains an end-to-end model with a unified loss that includes main and auxiliary tasks plus a contrastive term, achieving improvements in DSC, Jaccard, and connectivity (clDSC) on the 3D-IRCADb dataset, while noting some limitations in multi-class settings. The work highlights potential clinical impact by better capturing vascular hierarchy and suggests extensions to more scales, priors, memory banks for contrastive learning, and application to other vascular territories.

Abstract

Extracting hepatic vessels from abdominal images is of high interest for clinicians since it allows to divide the liver into functionally-independent Couinaud segments. In this respect, an automated liver blood vessel extraction is widely summoned. Despite the significant growth in performance of semantic segmentation methodologies, preserving the complex multi-scale geometry of main vessels and ramifications remains a major challenge. This paper provides a new deep supervised approach for vessel segmentation, with a strong focus on representations arising from the different scales inherent to the vascular tree geometry. In particular, we propose a new clustering technique to decompose the tree into various scale levels, from tiny to large vessels. Then, we extend standard 3D UNet to multi-task learning by incorporating scale-specific auxiliary tasks and contrastive learning to encourage the discrimination between scales in the shared representation. Promising results, depicted in several evaluation metrics, are revealed on the public 3D-IRCADb dataset.

Figures (5)

• Figure 1: Proposed pipeline for deep liver vessel segmentation using scale-specific auxiliary multi-task contrastive learning.
• Figure 2: (a) Visual aid to interpret notations related to multi-scale vessel clustering (Sect.\ref{['ssec:clustering']}). (b) Example of 3-scale vasculature clustering applied on a synthetic vasculature jassi2011vascusynth.
• Figure 3: Encoder-decoder architecture involved in our multi-task contrastive pipeline. PReLu stands for Parametric ReLU.
• Figure 4: Letter-value plot of branch radii $\hat{r}_{j}$$(\mathrm{mm})$ per ground truth volume $\pmb{y}_i$ from the 3D-IRCADb soler20103d dataset.
• Figure 5: Qualitative liver vessel segmentation results on CT scans from the 3D-IRCADb soler20103d dataset.