A combined system with convolutional neural networks and transformers for automated quantification of left ventricular ejection fraction from 2D echocardiographic images

• A novel and lightweight algorithm hybridizing CNN and transformer was developed for the measurement of LVEF. • The performance of the algorithm was validated in an external validation cohort. • The lightweight algorithm was equipped with a domestic echocardiography device. Accurate measurement of left ventricular ejection fraction (LVEF) is crucial in diagnosing and managing cardiac conditions. Deep learning (DL) models offer potential to improve the consistency and efficiency of these measurements, reducing reliance on operator expertise. The aim of this study was to develop an innovative software-hardware combined device, featuring a novel DL algorithm for the automated quantification of LVEF from 2D echocardiographic images. A dataset of 2,113 patients admitted to the Affiliated Hospital of Qingdao University between January and June 2023 was assembled and split into training and test groups. Another 500 patients from another campus were prospectively collected as external validation group. The age, sex, reason for echocardiography and the type of patients were collected. Following standardized protocol training by senior echocardiographers using domestic ultrasound equipment, apical four-chamber view images were labeled manually and utilized for training our deep learning framework. This system combined convolutional neural networks (CNN) with transformers for enhanced image recognition and analysis. Combined with the model that was named QHAutoEF, a ‘one-touch’ software module was developed and integrated into the echocardiography hardware, providing intuitive, real-time visualization of LVEF measurements. The device's performance was evaluated with metrics such as the Dice coefficient and Jaccard index, along with computational efficiency indicators. The dice index, intersection over union, size, floating point operations per second and calculation time were used to compare the performance of our model with alternative deep learning architectures. Bland-Altman analysis and the receiver operating characteristic (ROC) curve were used for validation of the accuracy of the model. The scatter plot was used to evaluate the consistency of the manual and automated results among subgroups. Patients from external validation group were older than those from training group ((60±14) years vs. (55±16) years, respectively, P < 0.001). The gender distribution among three groups were showed no statistical difference (43 % vs. 42 % vs. 50 %, respectively, P = 0.095). Significant differences were showed among patients with different type (all P < 0.001) and reason for echocardiography (all P <0.001 except for other reasons). QHAutoEF achieved a high Dice index (0.942 at end-diastole, 0.917 at end-systole) with a notably compact model size (10.2 MB) and low computational cost (93.86 G floating point operations (FLOPs)). It exhibited high consistency with expert manual measurements (intraclass correlation coefficient (ICC) =0.90 (0.89, 0.92), P < 0.001) and excellent capability to differentiate patients with LVEF ≥60 % from those with reduced function, yielding an area under the operation curve (AUC) of 0.92 (0.90–0.95). Subgroup analysis showed a good correlation between QHAutoEF results and manual results from experienced experts among patients of different types ( R = 0.93, 0.73, 0.92, respectively, P <0.001) and ages ( R = 0.92, 0.94, 0.89, 0.91, 0.81, respectively, P <0.001). Our software-hardware device offers an improved solution for the automated measurement of LVEF, demonstrating not only high accuracy and consistency with manual expert measurements but also practical adaptability for clinical settings. This device might potentially support clinicians and augment clinical decision. Central Illustration: Step 1: Three self-developed frameworks and four open established frameworks were horizontally compared in a public dataset; Step 2: The optimal framework was trained from scratch in a new Chinese dataset; Step 3: Validation of the model performance in an external Chinese cohort; Step 4:The device integration and one-touch trigger function; Step 5: The visualization results by one-touch operation.