Skip to content
Back to Portfolio
PulmoScore - COVID-19 severity scoring on chest X-rays

PulmoScore

Two-stage ML workflow for COVID-19 severity scoring on chest X-rays

Client: Beth Israel Deaconess Medical Center · Boston · United States 🇺🇸
PyTorchscikit-learnAlbumentationsSuperviselyWeights & BiasesOpenCV

Highlights

Situation

Beth Israel Deaconess Medical Center needed a faster, more accurate tool for diagnosing COVID-19 from chest X-rays.

Task

Design and implement an ML pipeline to segment lungs and assess disease severity on chest radiographs.

Action

Built a two-stage model using DeepLabV3+ and MA-Net, trained on diverse datasets for robust lung and infection detection.

Result

Achieved 0.30 MAE out of 6.0 and 11× faster processing, enabling scalable, real-time diagnostics.

Core Team

Viacheslav Danilov

Viacheslav Danilov

Lead ML Engineer

Quantori

Cambridge · United States 🇺🇸

Alex Proutski

Alex Proutski

Research Scientist

Quantori

Hague · Netherlands 🇳🇱

David Nefaridze

David Nefaridze

Data Scientist

Quantori

Tbilisi · Georgia 🇬🇪

Diana Litmanovich

Diana Litmanovich

Chief of Cardiothoracic Imaging

Beth Israel Deaconess Medical Center

Boston · United States 🇺🇸

Yuriy Gankin

Yuriy Gankin

Chief Science Officer

Quantori

Cambridge · United States 🇺🇸

Overview

This project developed a two-stage ML-driven workflow for segmenting and scoring lung diseases on chest X-rays, with a focus on COVID-19. The approach combined lung segmentation and disease localization, followed by a severity scoring process to provide quantitative assessments aligned with radiological practices.

Using publicly available datasets, the best-performing solution integrated DeepLabV3+ for lung segmentation and MA-Net for disease segmentation, achieving a mean absolute error (MAE) of 0.30 out of 6.0 – significantly outperforming established methods like BS-net and COVID-Net-S. The workflow is 11 times faster than comparable solutions and adaptable to broader applications, including pneumonia and tuberculosis.

Data

The dataset for this study was carefully curated to address both lung segmentation and disease-scoring tasks, drawing from diverse publicly available sources:

  • •Lung Segmentation: 6,810 X-rays from three datasets (Darwin, Montgomery, Shenzhen) providing diverse examples of lung pathologies and healthy subjects
  • •Disease Segmentation: 1,364 X-rays from four COVID-19 datasets (ACCD, CRD, CCXD, FCXD) and two normal datasets (CXN, RSNA) representing 580 COVID-19 and 784 normal cases
  • •Source Diversity: Data sourced from over 40 institutions worldwide, ensuring robust model generalization

Methods

The proposed workflow inherits the quantification and qualification of lung diseases from expert radiologists and fulfills the following processing steps (Figure 1):

  • Stage I (Lung Segmentation): Pixel-level localization of the lungs and removal of irrelevant areas using DeepLabV3+ architecture.
  • Stage II (Disease Segmentation): Identification of infected lung regions with nine tested neural networks: U-Net, U-Net++, DeepLabV3, DeepLabV3+, FPN, LinkNet, PSPNet, PAN, and MA-Net.
  • Stage III (Severity Scoring): Quantification and qualification of infected regions to compute an overall severity score per patient, aligned with radiological practices.
Figure 1. Two-stage COVID severity scoring workflow showing lung segmentation, disease detection, and severity quantification.

Results

The proposed workflow demonstrated high accuracy and efficiency in segmenting lungs, identifying infected regions, and computing severity scores for COVID-19 patients:

  • Best Accuracy: DeepLabV3+ (Stage I) + MA-Net (Stage II) achieved MAE of 0.30 out of 6.0, significantly outperforming BS-net (MAE: 2.52) and COVID-Net-S (MAE: 1.83).
  • Processing Speed: The fastest solution (DeepLabV3+ + PSPNet) processed 12.5 images/second, while the most accurate achieved 7.9 images/second.
  • Efficiency Gain: 11× faster than BS-net and COVID-Net-S (0.7 and 0.6 images/s respectively).

The model comparison chart (Figure 2) summarizes the performance across all tested architectures, while Figure 3 shows detailed segmentation results for each neural network.

Figure 2. Performance comparison of tested neural network architectures showing MAE and processing speed.

U-Net · Score: 4

U-Net++ · Score: 5

DeepLabV3 · Score: 0

DeepLabV3+ · Score: 3

FPN · Score: 4

LinkNet · Score: 3

PSPNet · Score: 3

PAN · Score: 5

MA-Net · Score: 5

Figure 3. Segmentation results comparison across all tested architectures. Cyan delineation: lung segmentation (Stage I); Red mask: disease prediction (Stage II); Yellow mask: ground truth.

Conclusion

The proposed two-stage workflow accurately segmented lung regions and quantified infection severity, offering a significant improvement over existing methods in both accuracy and speed. By combining DeepLabV3+ for lung segmentation and MA-Net for disease segmentation, the system provided reliable diagnostic support for COVID-19 and other respiratory diseases.

Future directions include adapting this workflow for real-time clinical deployment and expanding its application to other pulmonary conditions, such as pneumonia and tuberculosis.

Keep Exploring

Medical Imaging

Attention-guided deep learning for COVID-19 and pneumonia detection in chest X-rays, achieving 84% accuracy with Grad-CAM supervision.

January 2022
Image-Guided Surgery

Real-time landmark tracking for safer valve implantation during TAVI procedures using multi-task deep learning with 97% accuracy.

July 2021