Medical ultrasound has relied for decades on the speed of sound in tissues to produce clear images and accurate diagnoses. However, estimating this parameter from raw pulse-echo channel data is a nonlinear inverse problem that traditionally requires expensive labels or intensive computational simulations. Researchers have proposed IQ-JEPA, a self-supervised learning approach that operates directly on in-phase and quadrature (IQ) signals to predict sound speed without needing labels in most cases. This breakthrough promises to revolutionize quantitative ultrasound, drastically reducing dependence on labeled data and accelerating the adoption of artificial intelligence (AI) models in clinical settings.
The core of IQ-JEPA lies in an encoder based on a Hermitian transformer that processes the complex IQ signal, exploiting invariance to constant phase shifts. Sound speed manifests as a phase difference between regions, and the model learns to predict latent representations of masked regions from visible context, similar to how JEPA (Joint Embedding Predictive Architecture) works in computer vision. This architecture allows the encoder to be equivariant to phase and its conjugate-product feed-forward layer to be invariant, reading a quantity analogous to that used by classical coherence methods.
Experimental results are compelling: with 79,293 Fullwave 2.5 simulations at 2.5 MHz, pretraining on 63,435 unlabeled acquisitions achieves an error of 15.60 m/s with only 10,000 labels, a roughly threefold improvement in label efficiency over supervised training, growing to over fourfold at 1,000 labels. With the full label set, error drops to 8.71 m/s. Self-supervision is the dominant factor, and gains increase with more unlabeled data. Moreover, frozen encoder features reveal both sound speed and attenuation, and cross-domain transfer (e.g., between abdominal and layered phantoms) incurs little accuracy loss.
From a business perspective, this development represents an opportunity for companies specializing in custom software applications to integrate cutting-edge AI into medical products. Q2BSTUDIO, as a software and technology company, can help implement similar architectures in real environments, combining self-supervised pretraining with cloud solutions and cybersecurity to ensure patient data privacy. The ability to work with few labels is crucial in domains where annotation is expensive, such as ultrasound, and perfectly aligns with our AI services for automating diagnostic processes.
Integrating IQ-JEPA with cloud platforms like AWS or Azure allows scaling training to large volumes of unlabeled data, reducing infrastructure costs. Additionally, cybersecurity is essential to protect patient data during training and inference. Q2BSTUDIO offers tailored cybersecurity solutions ensuring compliance with regulations such as HIPAA or GDPR. We can also deploy Business Intelligence dashboards (Power BI) to visualize model performance metrics, or create AI agents that monitor ultrasound image quality in real time.
In the near future, models like IQ-JEPA could become the foundation for a quantitative ultrasound foundation model, similar to how large language models have transformed text processing. For this to be viable, companies need technology partners that understand both the clinical domain and software engineering. Q2BSTUDIO is positioned to bridge that gap by developing custom applications that integrate these innovations, whether on-premise or in the cloud, with security and scalability guarantees.
In conclusion, IQ-JEPA not only demonstrates that sound speed can be estimated with few labels, but also opens the door to a new generation of AI-based diagnostic tools. The combination of self-supervision, Hermitian transformers, and complex signal processing offers an efficient and accurate path forward for quantitative ultrasound. At Q2BSTUDIO, we are ready to help organizations leverage these techniques, offering comprehensive software development, AI, cloud, and cybersecurity services that transform research into real clinical products.
