Hybrid Approach for Automatic Annotation and Displacement Estimation of Cardiac Fluid Hemodynamics

In the present study, we have developed an innovative deep learning framework for per‐pixel estimation of cardiac hemodynamic displacements using Doppler echocardiography imaging coupled with electrocardiogram (ECG). Our hybrid proposed approach integrates an automatic annotation phase based on Farnebäck optical flow algorithm followed by a dense blood flow displacement estimation phase based on deep learning architectures. Two specific networks have been implemented: Echo‐FlowNetS‐1 and Echo‐FlowNetS‐2. The former integrates MaxPooling and dropout operations within the encoder and adopts a refined upsampling strategy in the decoder, yielding a 24.4% reduction in the computational complexity, while maintaining satisfactory results (AEE = 0.525 ± 0.157 px (0.234 ± 0.070 mm) [CI 95%: 0.215–0.252], Error_V ≈ 7.02 ± 2.10 mm/s[CI 95%: 6.45–7.59], RMSE = 2.095 ± 1.152 px (0.932 ± 0.513 mm)[CI 95%: 0.797–1.068], AAE = 1.563° ± 0.004°[CI 95%: 1.559–1.568] and Fl‐all = 4.74% ± 0.04%[CI 95%: 4.72–4.75]). The latter is an enriched version with additional post‐concatenation convolutional layers and bicubic upsampling in the decoder, leading to an enhancement in the accuracy of small displacements estimation (AEE = 0.482 ± 0.162 px (0.214 ± 0.072 mm)[CI 95%: 0.195–0.234], Error_V ≈ 6.42 ± 2.16 mm/s[CI 95%: 5.84–7.00], RMSE = 2.094 ± 1.151 px (0.931 ± 0.512 mm)[CI 95%: 0.797–1.068], AAE = 1.566° ± 0.037°[CI 95%: 1.556–1.575] and Fl‐all = 4.70% ± 0.04%[CI 95%: 4.69–4.71]), the experimental evaluation was conducted on our private cardiac ultrasound image database synchronized with ECG. The obtained results demonstrate that the proposed framework consistently outperforms existing reference methods, particularly in terms of robustness, density of the estimated flow field, and reproducibility across cardiac cycles.