In the present study, our results showed that, in patients eligible for GLS-based assessments, AI-based GLS was moderately correlated with conventional GLS. The agreements were consistent across abnormalities of cardiac structure and function (i.e., LV systolic function, LV hypertrophy, LA enlargement and diastolic dysfunction), and cardiovascular diseases/risks. Additionally, the agreement between GLS beginner and an experienced echocardiographer was numerically better in the AI-based method than the conventional method. These findings support the utility of AI-based GLS and enhance its application among patients undergoing for chemotherapies in clinical practice.

The incorporation of AI into cardiovascular image analysis software is ongoing [18]. Its fundamental applications include the assessment of cardiac function by LVEF and GLS [10, 18]. There are roughly two AI approaches for echocardiography-based GLS calculation. The first approach, such as used in the present study, utilizes deep learning to perform endocardial or full cardiac wall segmentation. This segmentation subsequently serves as input for a motion calculation algorithm, (e.g., by speckle tracking or optical flow methods), estimating the displacement and translation of the cardiac wall throughout the cardiac cycle [19, 20]. The second approach utilizes deep learning for both the segmentation and motion tracking steps [21].

To date, performance and validation data of these methods often comprise of head-to-head comparison between the AI-based GLS method and conventional-based method. Herein, correlation coefficients between the methods ranged from moderate to excellent ranging between 0.54 and 0.93, respectively [19,20,21,22]. For example, in a community-based cohort including 561 patients with a risk of developing heart failure, the agreement between manual and fully automated GLS showed fairness (rho = 0.69), and its agreement was improved with semi-automated GLS (rho = 0.84). However, the limited expertise in performing strain echocardiography may constrain the generalizability of these findings [22]. With this regard, our results showed numerically better reproducibility of AI-based GLS method than conventional GLS method. These findings may be explained by the fact that our AI approach can result in a hybrid approach combining AI for segmentation as initialization step for subsequent strain calculation, subsequently reducing the inter observer variability, and suggested this AI tool could be more user-friendly, particularly for unexperienced physicians.

Patients with abnormalities of LV structure and function were more likely to have lower value of GLS, compared with those not [23, 24]. Along with technical issues (i.e., tracing process, LV segment identification), different cardiac morphology and/or cardiovascular diseases may impact the agreement between AI-based GLS and conventional GLS [20]. Nonetheless, our results showed that the agreements between two GLS approaches were consistent across different abnormal cardiac structure, function and cardiovascular diseases/risks. These results are reassuring and may boost the applicability of our AI-based GLS approach.

Previously, the EACVI/ASE Strain Standardization Taskforce assessed the inter-vendor variability of nine conventional-based GLS method [25]. Future investigations by the Taskforce could reveal the performance and robustness characteristics of the different AI approaches for calculating GLS, i.e., based fully on deep learning, or a hybrid approach combining AI for segmentation as initialization step, and subsequent using the widely adopted speckle tracking or optical flow methods. Importantly, the merits of our AI-based GLS lie in assessing GLS values within pre-existing images through web-based software (eliminating the need for desktop installation) and in achieving better reproducibility of AI-based GLS across both inexperienced and experienced echocardiographers, compared to conventional GLS. Patients who are candidates for cancer treatment are recommended to undergo GLS measurements at least every three months during the treatment period [26, 27]. Therefore, our AI-based echocardiogram may hold potential for increasing the relevance of GLS assessments for patients treated with chemotherapies in clinical practice.

The results should be interpreted in light of the following limitations. This retrospective study has a small sample size, which may limit the generalizability of our findings and necessitate further statistical power to confirm them. AI-based GLS value was determined through assessment with good image quality, < 2 missing segments of the 6 individual myocardial segments in the 4-chamber view. All echocardiographic images were captured by experienced echocardiographic examiners; hence, any differences of generating the images due to varying levels of expertise were not accounted for in this analysis. Nevertheless, the utilization of an AI-based methodology might offer a solution to the challenge of precisely delineating the endocardium. The timing of performing echocardiography, particularly the duration after chemotherapy may impact patient cardiac function, however, we lacked a consistent duration period between chemotherapies and echocardiography assessment. Additionally, the longer training period for GLS beginners may influence our results. An adequately powered prospective study following structured training should be warranted.

Although GLS value may vary with vendors and examiners [25, 28], in the current analysis, GLS values, which experienced examiners measured using the conventional method, were considered as a standardized value. The influence of observer experience level on performance and the concordance of AI-based GLS values with those from different imaging modalities, such as cardiac magnetic resonance imaging, are sparsely reported and warrant further investigation [29].