Learning Dynamic Collaborative Network for Semi-supervised 3D Vessel Segmentation

TL;DR

The paper addresses semi-supervised 3D vessel segmentation under limited labeled data and complex vessel topologies. It proposes a dynamic collaborative network (DiCo) with two heterogeneous sub-networks (CNN and Transformer) that dynamically switch teacher and student roles based on labeled-data losses, complemented by a multi-view integration module and MIP-based adversarial supervision to enforce vessel-shape priors. Key contributions include dynamic role-switching to mitigate error propagation, a lightweight multi-view input strategy for local-global feature fusion, and 2D MIP adversarial supervision to shape predictions on unlabeled data. Experiments on three datasets show state-of-the-art semi-supervised performance with limited labels, highlighting DiCo’s practical impact for data-constrained clinical imaging.

Abstract

In this paper, we present a new dynamic collaborative network for semi-supervised 3D vessel segmentation, termed DiCo. Conventional mean teacher (MT) methods typically employ a static approach, where the roles of the teacher and student models are fixed. However, due to the complexity of 3D vessel data, the teacher model may not always outperform the student model, leading to cognitive biases that can limit performance. To address this issue, we propose a dynamic collaborative network that allows the two models to dynamically switch their teacher-student roles. Additionally, we introduce a multi-view integration module to capture various perspectives of the inputs, mirroring the way doctors conduct medical analysis. We also incorporate adversarial supervision to constrain the shape of the segmented vessels in unlabeled data. In this process, the 3D volume is projected into 2D views to mitigate the impact of label inconsistencies. Experiments demonstrate that our DiCo method sets new state-of-the-art performance on three 3D vessel segmentation benchmarks. The code repository address is https://github.com/xujiaommcome/DiCo

Figures (7)

• Figure 1: Semi-supervised segmentation frameworks. The dotted arrows indicate the supervision information flow of unlabeled data.
• Figure 2: Architecture of the proposed DiCo method. It consists of three fundamental components: the dynamic collaborative network, the multi-view integration module, and the MIP adversarial supervision module.
• Figure 3: Multi-view integration module.
• Figure 4: MIP adversarial supervision module.
• Figure 5: Visual segmentation examples from ImageCAS dataset.