Can we free echocardiographic videos from reverberation noise?

Doctors often struggle to read transthoracic echocardiogram videos because of reverberation clutter. A new AI‑based system can clean these images in real time. The network uses a 3D convolutional autoencoder that processes data across time and space. Two key features make it special: an attention mechanism that highlights noisy regions and uses surrounding context, and residual learning that preserves fine heart structures. To learn, the model was exposed to simulated artifact patterns overlaid on ultra‑realistic synthetic TTE sequences from six ultrasound vendors. Clean sequences served as ground truth. Tests on unseen real‑world data showed strong generalisation, even though training used only synthetic data. After filtering, segmental strain curves from noisy videos closely match those from clutter‑free originals. The system processes a full sequence in a fraction of a second, enabling real‑time filtering and potentially improving the accuracy of clinical indices derived from TTE. Source code and example videos are available on GitHub.

The proposed network is a 3D convolutional autoencoder operating on TTE image sequences, capturing the spatiotemporal nature of reverberation clutter throughout the cardiac cycle. 3D‑conv layers extract features from (time × space × space) volumes, suppressing noise that propagates over time. The attention module computes relevance weights for each pixel based on its surrounding context, allowing the network to focus on the most contaminated regions. Residual learning adds a direct skip connection between input and output, preserving fine structures such as valve motions. Training leverages a diverse set of simulated artifacts, generated by overlaying realistic patterns onto synthetic sequences from six ultrasound vendors, ensuring a rich variety of noise characteristics. Evaluation on unseen in‑vivo data confirms strong generalisation despite synthetic‑only training. Segmental strain analysis shows that after filtering, the discrepancy between noisy and clean segments is markedly reduced, indicating preservation of clinically meaningful information. Processing time under one second makes the system suitable for point‑of‑care use and real‑time integration into echocardiographic workflows.