Skip to content
Back to Portfolio
Histo Scanner - Microvascular feature segmentation in TEVG

Histo Scanner

Microvascular feature segmentation in tissue-engineered vascular grafts

Client: Kemerovo Cardiology Center · Kemerovo · Russia 🇷🇺
PyTorchPyTorch LightningAlbumentationsWeights & BiasesOpenSlideOpenCV

Highlights

Situation

Clinical biologists at the Kemerovo Cardiology Center needed accurate segmentation of microvascular features in TEVG histology images to assess vascular regeneration in sheep models.

Task

Lead model development and deliver a robust segmentation pipeline for nine histological features.

Action

Optimized six deep learning architectures and created an ensemble model using cross-validation, augmentation, and Bayesian tuning.

Result

Achieved a mean Dice score of 89%, enabling precise tissue analysis and accelerating biomedical research.

Core Team

Viacheslav Danilov

Viacheslav Danilov

Lead Data Scientist

Pompeu Fabra University

Barcelona · Spain 🇪🇸

Vladislav Laptev

Vladislav Laptev

Senior Data Scientist

Siberian Medical University

Tomsk · Russia 🇷🇺

Kirill Klyshnikov

Kirill Klyshnikov

Biomedical Scientist

Kemerovo Cardiology Center

Kemerovo · Russia 🇷🇺

Evgeny Ovcharenko

Evgeny Ovcharenko

Biomedical Engineer

Kemerovo Cardiology Center

Kemerovo · Russia 🇷🇺

Anton Kutikhin

Anton Kutikhin

Pathologist

Kemerovo Cardiology Center

Kemerovo · Russia 🇷🇺

Overview

This project focused on developing a deep learning-powered pipeline for segmenting and quantifying histological features in tissue-engineered vascular grafts (TEVGs). Using advanced deep learning models, the study aimed to address challenges in analyzing complex tissue regeneration patterns within biodegradable grafts implanted in sheep carotid arteries.

By optimizing six neural networks and creating an ensemble model, the project achieved precise segmentation of key microvascular features, including arterioles, venules, capillaries, immune cells, and nerve trunks. The ensemble model demonstrated state-of-the-art accuracy, achieving a mean Dice Similarity Coefficient (DSC) of 0.889, offering a robust tool for accelerating research in tissue engineering and translational medicine.

Data

The study utilized a dataset comprising 104 Whole Slide Images (WSIs) obtained from biodegradable TEVGs implanted into the carotid arteries of 20 sheep. After six months, the sheep were euthanized to assess vascular tissue regeneration patterns.

  • •Patches: WSIs were automatically sliced into 99,831 patches, of which 1,401 were annotated by pathologists (Figure 1).
  • •Features: Nine histological features: arteriole lumen (AL), arteriole media (AM), arteriole adventitia (AA), venule lumen (VL), venule wall (VW), capillary lumen (CL), capillary wall (CW), immune cells (IC), and nerve trunks (NT).
  • •Verification: A lead pathologist verified all annotations to ensure high-quality ground truth data for model training and evaluation.
Figure 1. Annotation methodology showing the patch generation and manual labeling workflow for nine histological features.

Methods

The methodology was designed to address the complexity of segmenting microvascular features in tissue-engineered vascular grafts by combining advanced deep learning techniques with rigorous evaluation protocols:

  • Patch Generation: Whole slide images were divided into smaller, high-resolution patches to enable efficient processing.
  • Model Optimization: Six deep learning models (U-Net, MA-Net, DeepLabV3, PSPNet, FPN, LinkNet) were optimized using Bayesian hyperparameter tuning and HyperBand early stopping strategies.
  • Cross-Validation: A 5-fold cross-validation approach was applied to ensure robust evaluation while preventing data leakage.
  • Augmentation Techniques: Training datasets were enhanced using augmentation methods, including rotations, flips, and brightness adjustments, to improve generalization and mitigate overfitting.
  • Ensemble Model: Predictions from the top-performing models were combined into an ensemble model to maximize segmentation accuracy.

Results

The results demonstrate the effectiveness of the proposed pipeline in accurately segmenting microvascular features of TEVGs. Individual models performed well across various features, while the ensemble model significantly improved overall accuracy:

  • •Best Individual: MA-Net achieved the highest individual performance with a mean DSC of 0.875, excelling in arteriole segmentation (Figure 2).
  • •Ensemble Model: Outperformed all individual networks, achieving a mean DSC of 0.889, with significant improvements in venule, capillary, and immune cell segmentation.
  • •Feature Accuracy: High precision for critical elements like arteriole lumen (DSC: 0.939) and nerve trunks (DSC: 0.978).
Figure 2. Performance comparison of all models, highlighting MA-Net's superior results.

The ensemble of three models (MA-Net, DeepLabV3, and FPN) provided the optimal solution (Figure 3).

Figure 3. Example patches showcasing the segmentation of histologic features using the ensemble model.

Conclusion

This project demonstrated the effectiveness of deep learning models in segmenting microvascular structures within tissue-engineered vascular grafts. By integrating advanced algorithms like MA-Net and DeepLabV3, the workflow achieved high precision in analyzing tissue regeneration patterns.

These findings pave the way for enhanced automation in tissue engineering research. Future work could explore extending this pipeline to analyze more complex vascular networks or incorporating foundation models for broader applicability.

Keep Exploring

Biomedical Research

AI-powered microscopy pipeline for detecting and counting cancer cell biomarkers using EfficientDet, achieving 85% mAP for nuclei detection.

January 2021
Medical Imaging

Explainable AI framework for detecting pulmonary edema features in chest X-rays using deep learning segmentation and object detection.

March 2024