VPBSD:Vessel-Pattern-Based Semi-Supervised Distillation for Efficient 3D Microscopic Cerebrovascular Segmentation

TL;DR

Experimental results on real-world data, including comparisons with state-of-the-art methods and ablation studies, demonstrate that the proposed VpbSD pipeline and its individual components effectively address the challenges inherent in microscopic cerebrovascular segmentation.

Abstract

3D microscopic cerebrovascular images are characterized by their high resolution, presenting significant annotation challenges, large data volumes, and intricate variations in detail. Together, these factors make achieving high-quality, efficient whole-brain segmentation particularly demanding. In this paper, we propose a novel Vessel-Pattern-Based Semi-Supervised Distillation pipeline (VpbSD) to address the challenges of 3D microscopic cerebrovascular segmentation. This pipeline initially constructs a vessel-pattern codebook that captures diverse vascular structures from unlabeled data during the teacher model's pretraining phase. In the knowledge distillation stage, the codebook facilitates the transfer of rich knowledge from a heterogeneous teacher model to a student model, while the semi-supervised approach further enhances the student model's exposure to diverse learning samples. Experimental results on real-world data, including comparisons with state-of-the-art methods and ablation studies, demonstrate that our pipeline and its individual components effectively address the challenges inherent in microscopic cerebrovascular segmentation.

Figures (4)

• Figure 1: A set of 2D segmentation results selected from the final 3D reconstruction of the whole brain segmentation for a CD1 mouse strain.
• Figure 2: The training workflow of VpbSD. The black arrows indicate the direction of forward propagation, while the red arrows represent regulatory signals. During the self-supervised pretraining phase of the teacher model, the unlabeled microscopic (MS) volume$X$ is input into the model, which trains by computing the codebook loss $L_{cb}^T$ and reconstruction loss $L_{rec}^T$. In the fine-tuning phase,labeled volumes are employed to calculate the fully supervised segmentation loss $L_{seg}^T$. In the knowledge distillation phase, the full supervision segmentation loss $L_{seg}^S$ is computed for labeled data, while a semi-supervised loss $L_{semi}^S$ is calculated for unlabeled data using pseudo ground truth from the teacher model and predictions from the student model. Additionally, a distillation loss $L_{dis}^S$ is computed between the quantized encoded results from the teacher model and the encoded results from the student model.
• Figure 3: An illustration of the vessel-pattern-based knowledge distillation module.
• Figure 4: Visual comparative segmentation results between the proposed methods and existing approaches. Each example presents the input image alongside the ground truth (label) and the corresponding segmentation outcomes. Green regions denote accurate segmentations, blue regions indicate missed detections, and red regions signify over-segmentations. Two representative examples are provided for thorough evaluation.