Arrhythmias & electrophysiology
Deep learning estimation of three-dimensional left atrial shape from two-chamber and four-chamber cardiac long axis views
Authors' conclusion
Compared to the bi-plane area-length method, the network showed higher accuracy and robustness for both volume and surface area.
Study context
- Population
- Coronary computed tomography angiography (CCTA) cases (training/validation/testing) and an independent cardiac magnetic resonance (CMR) test set from another institutionplus independent CMR test set n = 20
- Participants
- 1,700 cases
- Comparison
- Network vs Area-length method
- Outcome
- LA volume / surface area estimation, CCTA test set (200 cases)
Methodological detailsexternal validation in independent dataset from another institution, CCTA cases split 1400/100/200 for training/validating/testing
- External validation in independent dataset from another institution
- CCTA cases split 1400/100/200 for training/validating/testing
Performance
As reported in the abstract
| Test or model | Mean Dice score | Mean absolute error, LA volume | Mean absolute error, LA surface area |
|---|---|---|---|
| Network | 93.7% | 3.5 mL | 4.9 cm² |
| Area-length method | 13.0 mL | 34.1 cm² |
Mean Dice score
Network93.7%
Area-length methodNot reported in the abstract
Mean absolute error, LA volume
Network3.5 mL
Area-length method13.0 mL
Mean absolute error, LA surface area
Network4.9 cm²
Area-length method34.1 cm²
- Network vs Area-length method: P < 0.05 — lower mean absolute error for both volume and surface area
- 3D segmentation Dice (comparator for shape reconstruction)
- 97.4%
Further findings
| Outcome | Network | Area-length method | P value | |
|---|---|---|---|---|
| LA shape, volume and surface area, independent CMR test setLower mean absolute error for both volume and surface area | P < 0.05 | |||
| Mean Dice score | 87.4% | Not reported in the abstract | Area-length method: not reported in the abstract | |
| Mean absolute error, LA volume | 6.0 mL | 14.2 mL | ||
| Mean absolute error, LA surface area | 5.7 cm² | 19.3 cm² |
- LA shape, volume and surface area, independent CMR test set: n = 20
Original publication
Deep learning estimation of three-dimensional left atrial shape from two-chamber and four-chamber cardiac long axis views. Eur Heart J Cardiovasc Imaging. Published 2 February 2023. doi:10.
Summarised from the published abstract; every number is checked against it