tUbeNet: a generalizable deep learning tool for 3D vessel segmentation

Area of Science:

• * Medical Imaging Analysis
• * Computational Biology
• * Deep Learning Applications

Background:

• * Existing cell annotation software like Cellpose is widely used in bioimage analysis.
• * There is a lack of equivalent tools for three-dimensional (3D) vascular annotation.
• * The vascular system's involvement in various diseases necessitates quantitative analysis of vascular imaging.

Purpose of the Study:

• * To develop a generalizable deep learning model for 3D vascular segmentation.
• * To create a human-in-the-loop training approach for efficient model fine-tuning.
• * To enable accurate 3D vascular annotation across different tissues, modalities, scales, and pathologies.

Main Methods:

• * A 3D convolutional neural network was trained on diverse imaging modalities (optical, CT, photoacoustic).
• * A pre-trained 'foundation' model was fine-tuned using minimal manually labeled ground truth data.
• * The model learned common vascular features across modalities and scales through varied training data.

Main Results:

• * The foundation model specialized to new datasets with as little as 0.3% of the volume for fine-tuning.
• * Segmentations achieved high accuracy (DICE coefficient 0.81–0.98) across various applications.
• * The model demonstrated generalizability across tissues, modalities, scales, and pathologies.

Conclusions:

• * A generalizable deep learning model for 3D vascular segmentation can be effectively specialized with minimal human input.
• * This approach significantly reduces the need for extensive manual annotation of training data.
• * The developed model and training strategy facilitate accurate 3D vascular network segmentation for medical research.