CT Pneumonia Detection System

A deep learning model for accurate detection of COVID-19 and differentiation from community acquired pneumonia and other lung diseases.

The following demonstration is intended to illustrate an application of AI in radiology and to showcase the capabilities of the Keya Medical team. This product is not for sale in the USA.

COVID-19 Research Results Published in Radiology

On March 20, the research results of Keya Medical’s COVID-19 AI Assisted Diagnosis algorithm were published in the peer-reviewed journal Radiology.
RADIOLOGY ARTICLEDEEPVESSEL BLOG

CT Pneumonia Detection System

CT Pneumonia Detection System aims to accurately detect COVID-19 and differentiate it from Community Acquired Pneumonia (CAP) and other lung diseases.

Powered by cutting-edge artificial intelligence and medical image analysis technologies, CT Pneumonia Detection System helps radiologists diagnose and assess pneumonia progress efficiently.

Worklist – Triage and Notification

Real-time lung CT images analysis

Automated detection of COVID-19 and differentiation from other types of pneumonia

Radiologists can choose their review priorities based on notifications on the worklist

Improve the efficiency of detecting COVID-19 or other pneumonia suspected patients

Image Analysis

Overview

Detection Results

  • Suspected COVID-19
  • Suspected Pneumonia
  • No inflammation identified

QA in Lung Lobes

  • Schematic diagram of infected regions
  • Quantitative assessment of lesion volume and lesion proportion

Lesion Histogram

  • Hounsfield unit (HU) distribution within lesion regions
  • Lesion volume and proportion grouped by different HU ranges

Findings

  • Summary of findings in each lobe for direct use: lesion volume, proportion, and characteristics (GGO, consolidation)

Lesion Segmentation

Automated lesion contouring for lesion identification and quantitative analysis
Multiplanar observation (Transverse/Coronal/Sagittal)
Draggable scrollbar for fast lesion localization

Follow-Up Comparison

Comparing two consecutive scans of the same patient quantitatively to assess the progress of pneumonia
Linking slices of the lungs from two consecutive scans for convenient observation and comparison
Changes in lesion proportions of each lobe are compared and shown using arrows (showing increased/decreased trend)

Validation

Clinical Datasets

Retrospective multi-center clinical validation

Hospitals

CAP and non-pneumonia patients were randomly selected (Aug 16, 2016 – Feb 17, 2020)

Patients

Avg Age: 49
Male: 1838
Female:1484

CT Scans

COVID-19: 1296
CAP: 1735
Other: 1325

COVID-19 Scans

Acquired between Dec 31, 2019, and Feb 17, 2020
All were confirmed by RT-PCR

Workflow & Methods

COVID-19 detection neural network – COVNet: based on deep neural network, AI-powered

  • Input: CT DICOM (Non-contrast CT, Slice thickness<=3mm)
  • Output: Disease Classification Labels (COVID-19, CAP, Non-Pneumonia)

Detection Performance

COVID-19 Detection:

Sensitivity 90%; Specificity 96%; AUC 0.96

CAP Detection:

Sensitivity 87%; Specificity 92%; AUC 0.95

We invite qualified medical centers to contact us to discuss how these capabilities could help with your response to the COVID-19 pandemic.

8 + 4 =