Calcium scoring in low-dose ungated chest CT scans using convolutional long-short term memory networks

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Cardiology

Background:

• Coronary artery calcium (CAC) scoring is crucial for cardiovascular risk assessment.
• Current methods for CAC quantification can be time-consuming and resource-intensive.
• Low-dose computed tomography attenuation correction maps (CTAC) offer potential for efficient CAC assessment.

Purpose of the Study:

• To develop and validate a novel deep learning method for automatic CAC quantification.
• To utilize a convolutional long-short-term memory (conv-LSTM) deep neural network for CAC scoring.
• To compare the performance of the conv-LSTM model against a U-Net reference model using CTAC scans.

Main Methods:

• A conv-LSTM model was trained on 9543 scans for CAC segmentation.
• A U-Net model was trained as a reference for comparison.
• Both models were validated on the OrCaCs dataset (n=32) and a held-out cohort (n=507).
• Agreement was assessed using Cohen's kappa coefficients and concordance matrices across four CAC score categories.

Main Results:

• The conv-LSTM model demonstrated significantly shorter median inference times (6.18s) compared to U-Net (10.1s) on a CPU (p<0.0001).
• Conv-LSTM exhibited substantially lower memory consumption during training (13.11 Gb) versus U-Net (22.31 Gb).
• Both models showed comparable agreement with expert annotations for CAC scoring.

Conclusions:

• The developed conv-LSTM method provides accurate CAC quantification from low-dose CTAC scans.
• This novel deep learning approach offers significant advantages in terms of speed and computational efficiency.
• Conv-LSTM represents a promising tool for improving the workflow of CAC scoring in clinical practice.