Unsupervised Microvascular Image Segmentation Using an Active Contours Mimicking Neural Network

Shir Gur, Lior Wolf, Lior Golgher, Pablo BlinderICCV 2019 · OpenReview

Abstract

The task of blood vessel segmentation in microscopy images is crucial for many diagnostic and research applications. However, vessels can look vastly different, depending on the transient imaging conditions, and collecting data for supervised training is laborious. We present a novel deep learning method for unsupervised segmentation of blood vessels. The method is inspired by the field of active contours and we introduce a new loss term, which is based on the morphological Active Contours Without Edges (ACWE) optimization method. The role of the morphological operators is played by novel pooling layers that are incorporated to the network's architecture. We demonstrate the challenges that are faced by previous supervised learning solutions, when the imaging conditions shift. Our unsupervised method is able to outperform such previous methods in both the labeled dataset, and when applied to similar but different datasets. Our code, as well as efficient PyTorch reimplementations of the baseline methods VesselNN and DeepVess is available on GitHub - https://github.com/shirgur/UMIS.

Introduction

Microscopic imaging of blood vessels, particularly through techniques like multi-photon microscopy, is essential for studying neuro-vascular interactions in vivo. Analyzing the resulting volumetric data requires fast and accurate 3D segmentation. Traditionally, supervised deep learning has been the standard approach. However, it suffers from two major drawbacks: the extreme difficulty of manually annotating 3D microvascular structures and poor generalization due to domain shifts in imaging conditions (e.g., varying signal-to-noise ratios and contrast). This paper introduces an unsupervised segmentation algorithm that circumvents the need for human labels, enabling better adaptation to diverse imaging environments.

Methodology

The core of the proposed method is a 3D ResNet34-based encoder-decoder network trained without ground-truth labels. The innovation lies in bridging classical computer vision with deep learning through a loss function that mimics Morphological Active Contours Without Edges (ACWE).

Architecture and Loss Functions

The network architecture includes a segmentation decoder ($D_{seg}$) and an auxiliary reconstruction decoder ($D_{rec}$), as illustrated in the following diagram:

Overview of the proposed architecture, featuring an encoder-decoder structure with auxiliary reconstruction and segmentation branches, and integrated morphological pooling layers.

Key components include:

  • Active Contour Loss ($L_{AC}$): Derived from the ACWE energy functional, this loss encourages the network to partition the image into vessel and background regions based on mean intensities, effectively evolving a boundary around bright structures.
  • Morphological Pooling Layers: To ensure spatial consistency and smoothness—tasks typically handled by partial differential equations in classical methods—the authors introduce differentiable, non-learned pooling layers. These layers use fixed $3 \times 3 \times 3$ structuring elements to perform morphological operations like dilation and erosion.
  • Auxiliary Losses: A reconstruction loss ensures the network preserves input information, while variance and entropy losses prevent the trivial solution of segmenting the entire image as a single class.

Datasets

The method was evaluated on three two-photon microscopy datasets:

  1. DeepVess Dataset: A 3D in-vivo vascular volume with expert annotations.
  2. VesselNN Dataset: 12 annotated 3D volumes of mouse cortex and tumor vasculature.
  3. 4D-NVIV Dataset: A new, unannotated 4D movie dataset tracking neurovascular activity, characterized by a larger field-of-view and lower SNR.

Results and Discussion

The unsupervised method demonstrates superior performance both within and across datasets, particularly when imaging conditions vary.

Qualitative comparison of segmentation results on DeepVess and VesselNN datasets across various methods, highlighting the robustness of the unsupervised approach.

Performance Metrics

On the DeepVess dataset, the unsupervised model achieved an F1-score of 0.829, surpassing the supervised DeepVess baseline (0.820). The advantages become even more pronounced in cross-dataset scenarios. When trained on VesselNN and tested on DeepVess, the unsupervised model achieved an Average Precision (AP) of 0.752, compared to 0.664 for the supervised model. Conversely, when trained on DeepVess and tested on VesselNN, it reached an AP of 0.902, significantly outperforming the supervised baseline's 0.701.

Qualitative Analysis

On the noisy 4D-NVIV dataset, the unsupervised approach successfully tracked penetrating vessels that caused supervised methods to fail or produce fragmented results. This highlights the power of transductive learning, where the model can be fine-tuned directly on unlabelled test data to adapt to specific imaging conditions.

Conclusion

By integrating classical mathematical principles into a neural network framework, this work provides a robust solution for unsupervised microvascular segmentation. The method's ability to generalize across domains and adapt at test-time makes it a valuable tool for high-throughput analysis of cerebral blood flow. Future work may involve refining the compound loss function's hyperparameters for broader application in other medical imaging modalities.

Summary updated

Abstract

The task of blood vessel segmentation in microscopy images is crucial for many diagnostic and research applications. However, vessels can look vastly different, depending on the transient imaging conditions, and collecting data for supervised training is laborious. We present a novel deep learning method for unsupervised segmentation of blood vessels. The method is inspired by the field of active contours and we introduce a new loss term, which is based on the morphological Active Contours Without Edges (ACWE) optimization method. The role of the morphological operators is played by novel pooling layers that are incorporated to the network's architecture. We demonstrate the challenges that are faced by previous supervised learning solutions, when the imaging conditions shift. Our unsupervised method is able to outperform such previous methods in both the labeled dataset, and when applied to similar but different datasets. Our code, as well as efficient pytorch reimplementations of the baseline methods VesselNN and DeepVess are attached as supplementary.
BibTeX
@inproceedings{gur2019unsupervised,
  author = {Shir Gur and Lior Wolf and Lior Golgher and Pablo Blinder},
  title = {{Unsupervised Microvascular Image Segmentation Using an Active Contours Mimicking Neural Network}},
  booktitle = {IEEE/CVF International Conference on Computer Vision},
  year = {2019},
  url = {https://openreview.net/forum?id=UdDsWFw-Bvc},
}
Figure