THE APPLICATION OF ARTIFICIAL INTELLIGENCE ON SCAR SEGMENTATION IN CARDIAC MAGNETIC RESONANCE IMAGING: A SYSTEMATIC LITERATURE REVIEW

OBJECTIVE We aimed to assess the application of Artificial Intelligence (AI) methodologies for scar segmentation in cardiac magnetic resonance (CMR) imaging and its performance evaluation. MATERIALS & METHODS Following PRISMA (Preferred Reporting Items for Systematic Review and Meta-Analyses) guidelines, a systematic search of PubMed and Science Direct was undertaken from 2012 to 2022 to search for full-text publications that implemented AI methods on scar segmentation in CMR in patients with cardiovascular diseases. RESULTS A total of 21 articles out of 475 articles were selected for the final review. Supervised deep learning and unsupervised machine learning were implemented in 16 (76.2%) and 5 (23.8%) articles respectively, favouring learning methods. Dice similarity coefficient (DSC) value was used as measures of the performance of AI methods in 19 articles. Supervised and unsupervised learning models had similar DSC compared to manual segmentation with a score of 0.74, 95% confidence interval (CI) [0.67, 0.81] vs 0.71, 95% CI [0.63, 0.79], P = 0.35). The application of AI has been advanced with the emerging of sophisticated algorithms allowing for quantification of border zone and microvascular obstruction regions. The performance of AI method is highly depending on the network architecture, training strategies, and data set used for training. CONCLUSION The presence of AI methods in scar segmentation demonstrated high feasibility with good performance evaluation for quantifying myocardial scar. This study can have a huge impact on clinicians in health care by improving their experiences with scar segmentation and enhancing clinically validated application of AI in CMR imaging. We aimed to assess the application of Artificial Intelligence (AI) methodologies for scar segmentation in cardiac magnetic resonance (CMR) imaging and its performance evaluation. Following PRISMA (Preferred Reporting Items for Systematic Review and Meta-Analyses) guidelines, a systematic search of PubMed and Science Direct was undertaken from 2012 to 2022 to search for full-text publications that implemented AI methods on scar segmentation in CMR in patients with cardiovascular diseases. A total of 21 articles out of 475 articles were selected for the final review. Supervised deep learning and unsupervised machine learning were implemented in 16 (76.2%) and 5 (23.8%) articles respectively, favouring learning methods. Dice similarity coefficient (DSC) value was used as measures of the performance of AI methods in 19 articles. Supervised and unsupervised learning models had similar DSC compared to manual segmentation with a score of 0.74, 95% confidence interval (CI) [0.67, 0.81] vs 0.71, 95% CI [0.63, 0.79], P = 0.35). The application of AI has been advanced with the emerging of sophisticated algorithms allowing for quantification of border zone and microvascular obstruction regions. The performance of AI method is highly depending on the network architecture, training strategies, and data set used for training. The presence of AI methods in scar segmentation demonstrated high feasibility with good performance evaluation for quantifying myocardial scar. This study can have a huge impact on clinicians in health care by improving their experiences with scar segmentation and enhancing clinically validated application of AI in CMR imaging.