Skip to content
Back to Portfolio
Stenosis Spotter - Automated coronary artery stenosis detection

Stenosis Spotter

Real-time coronary stenosis detection using deep learning for intraoperative decision support

Client: Kemerovo Cardiology Center · Kemerovo · Russia 🇷🇺
TensorFlowPythonAWS SageMakerLabelBoxscikit-learnOpenCV

Highlights

Situation

Cardiologists at the Kemerovo Cardiology Center needed accurate, real-time detection of coronary stenosis during surgery to reduce diagnostic delays.

Task

Lead deep learning model development and integration for intraoperative angiographic image analysis.

Action

Trained and evaluated 8 CNN detectors on 8,325 images, optimizing accuracy vs. inference speed using SageMaker and Bayesian tuning.

Result

RFCN ResNet-101 was deployed in live surgeries, improving diagnostic speed and supporting critical treatment decisions.

Core Team

Viacheslav Danilov

Viacheslav Danilov

Senior Data Scientist

University of Leeds

Leeds · United Kingdom 🇬🇧

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

Research Scientist

Kemerovo Cardiology Center

Kemerovo · Russia 🇷🇺

Alejandro Frangi

Alejandro Frangi

Professor

University of Leeds

Leeds · United Kingdom 🇬🇧

Overview

This project addresses a critical need in modern cardiology: fast and accurate detection of coronary artery stenosis during invasive angiography. Despite angiography being the gold standard for diagnosing coronary artery disease, its effectiveness can be limited by image noise, patient-specific anatomy, and operator variability. To reduce diagnostic errors and decision latency, this research implemented deep learning-based object detection models to identify stenotic lesions in real time.

Eight convolutional neural network architectures were trained and benchmarked on angiographic images from 100 patients. The study compared trade-offs between model accuracy and inference speed, critical for clinical deployment. Three models stood out: Faster-RCNN Inception ResNet (highest accuracy), SSD MobileNet (fastest), and RFCN ResNet-101 (best accuracy-speed balance). This work demonstrates that deep learning can be reliably integrated into real-time clinical workflows, improving diagnosis, treatment planning, and potentially surgical outcomes.

Data

The dataset consisted of 8,325 grayscale angiographic images (512×512 to 1000×1000 px) from 100 patients diagnosed with coronary artery disease:

  • •Acquisition: Images acquired using Coroscop (Siemens) and Innova (GE Healthcare) systems at the Kemerovo Cardiology Center
  • •Patient Criteria: All patients had either angiographically confirmed ≥70% stenosis or functionally verified ischemia
  • •Data Split: Training (80%), validation (10%), and test (10%) sets split by patient to preserve clinical independence
  • •Annotation: Manual labeling performed by trained cardiologists using the LabelBox platform

Methods

Eight CNN object detection models were evaluated to find the optimal balance between accuracy and real-time performance:

  • Model Architectures: SSD MobileNet V1/V2, SSD ResNet-50, Faster-RCNN ResNet-50/101, RFCN ResNet-101, Faster-RCNN Inception ResNet, and NASNet were tested for their complementary strengths.
  • Transfer Learning: All models were pretrained on the COCO dataset and fine-tuned on the medical angiographic dataset.
  • Training Infrastructure: Amazon SageMaker with Bayesian optimization for hyperparameter tuning across encoder type, input size, optimizer, and learning rate.
  • Loss Functions: Smooth L1 for localization and Focal Loss for classification to handle class imbalance effectively.
  • Optimization Strategies: Hard example mining (used in MobileNet V2) and architectural complexity analysis to understand accuracy-speed trade-offs.

Results

The comparative evaluation of eight deep learning models revealed three standout performers, each excelling in different aspects (Figure 1):

  • Faster-RCNN Inception ResNet (highest accuracy): mAP 0.95, F1-score 0.96, inference speed 3 FPS. Ideal for scenarios where precision is paramount, such as complex or borderline stenotic cases.
  • SSD MobileNet V2 (fastest): mAP 0.83, F1-score 0.80, inference speed 38 FPS. Well-suited for real-time use in high-throughput settings or emergency diagnostics.
  • RFCN ResNet-101 (best balance): mAP 0.94, F1-score 0.96, inference speed 10 FPS. Particularly well-suited for live clinical workflows, offering a practical compromise for quasi-real-time surgical environments.
Figure 1. Performance comparison of eight CNN object detection models on coronary angiography images.

The detection capabilities were further validated using angiographic sequences during minimally invasive surgeries (Figure 2), highlighting their clinical robustness and practical relevance.

Figure 2. Real-time stenosis detection during live minimally invasive surgery using RFCN ResNet-101.

Conclusion

This project demonstrated the feasibility of real-time coronary stenosis detection using deep learning. By identifying optimal trade-offs between inference speed and accuracy, the solution empowers interventional cardiologists with immediate, reliable decision support. The RFCN ResNet-101 model shows strong potential for real-world integration.

Future directions include enhancing multivessel disease detection, expanding dataset diversity, and integrating guidance systems for surgical intervention planning.

Keep Exploring

Image-Guided Surgery

Deep learning solution for surgical tool segmentation in 3D ultrasound, achieving 93.6% Dice score for real-time catheter localization.

February 2023
Biomedical Research

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

January 2021