Abstracts

RETURN TO ABSTRACT LISTING


2348. Validation of a Deep Neural Network-Based Fully Automatic Coronary Calcium Scoring Algorithm on Chest CT
Authors * Denotes Presenting Author
  1. Vincent Giovagnoli *; Medical University of South Carolina
  2. Marly van Assen; Medical University of South Carolina
  3. Simon Martin; Medical University of South Carolina
  4. Tilman Emrich; Medical University of South Carolina
  5. Richard Bayer; Medical University of South Carolina
  6. Akos Varga-Szemes; Medical University of South Carolina
  7. U. Joseph Schoepf; Medical University of South Carolina
Objective:
To evaluate deep-learning based automated calcium quantification on chest CTs compared with manual qualitative and quantitative assessment and standard Agatston score.

Materials and Methods:
A total of 95 consecutive patients with both dedicated cardiac and chest acquisitions were included in this study. The median number of days between the two CTs was 186 [76-383]. Additionally, 168 patients with chest CT only were included for the comparison with qualitative calcium classification. Automated calcium quantification was performed using a deep-learning model which combines a convolution neural network for image features and a fully connected neural network for spatial coordinate features. An additional CNN based aortic segmentation model is used to eliminate false positive aortic detections. The algorithm was trained on 1261 dedicated cardiac-CTs and subsequently refined on 500 non-gated chest-CTs. Results from the AI model were compared to Agatston-scores (cardiac-CTs), manually determined calcium volume (chest-CTs) and clinical classifications (no, mild, moderate and severe calcium).

Results:
The Agatston score and AI determined calcium volume showed high correlation (r= 0.921 (p<0.001) and R2=0.91). According to the Agatston risk categories, a total of 67 (70%) cases were correctly classified. Manual annotation of calcium volume (chest CT) of all 95 patients showed excellent correlation with the AI determined volumes (r=0.923 (p<0.001) and R2=0.96) and no significant differences (p=0.247). Based on the clinical qualitative analysis, 138 (82%) cases were correctly classified with a kappa coefficient of 0.74 representing good agreement. All wrongly classified scans (30 (18%)) were attributed to an adjacent category.

Conclusion:
Deep-learning-based calcium quantification on chest-CTs shows good correlation compared to both quantitative and qualitative reference standards. Fully automating this process may reduce evaluation time and optimize clinical calcium scoring without additional acquisitions. AI based calcium quantification on chest CTs, might aid in the evaluation of cardiovascular risk, especially in a screening setting. Using a fully automated approach can help with the increasing workload.