IQ-JEPA: A Joint-Embedding Predictive Architecture with a Hermitian Vision Transformer for Sound Speed and Attenuation Estimation from Ultrasound IQ Data

2026-07-24 · Source: Machine Learning · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Medical Imaging AI · Depth: Expert, quick

Summary

IQ-JEPA is a novel Joint-Embedding Predictive Architecture designed to estimate sound speed and attenuation from ultrasound in-phase and quadrature (IQ) data, addressing the nonlinear inverse problem of recovering sound speed from raw pulse-echo channel data. This architecture leverages both unlabeled and labeled data, employing a Hermitian vision transformer as its encoder. The transformer operates directly on complex signals, ensuring phase equivariance and invariance to constant phase offsets, mirroring classical coherence methods. Pretraining on 63,435 unlabeled acquisitions from 79,293 Fullwave 2.5 simulations at 2.5 MHz achieved a 15.60 m/s error with only 10,000 labels. This demonstrates a threefold gain in label efficiency compared to supervised training, escalating to over fourfold at 1,000 labels. While 2.2x below an InversionNet baseline and 8.71 m/s at full labels, the efficiency gain grows with more unlabeled pretraining data, highlighting the dominance of self-supervision. The system's ability to transfer frozen features and perform cross-distribution pretraining positions it as a foundational step for quantitative ultrasound models.

Key takeaway

For Machine Learning Engineers developing quantitative ultrasound models, IQ-JEPA offers a path to significantly reduce label dependency. You should explore self-supervised pretraining with large unlabeled datasets to achieve substantial label efficiency gains, potentially threefold or more. Consider implementing Hermitian Vision Transformers for direct complex signal processing; this improves accuracy and transferability across phantoms. This method could accelerate the deployment of robust diagnostic tools.

Key insights

IQ-JEPA uses a Hermitian Vision Transformer and self-supervision to efficiently estimate ultrasound sound speed and attenuation from IQ data.

Principles

• Self-supervision significantly boosts label efficiency in inverse problems.
• Phase-equivariant transformers can process complex signals directly.
• Unlabeled data pretraining improves model transferability.

Method

IQ-JEPA pretrains an encoder on masked IQ regions to predict latent representations from visible context, then fine-tunes on simulated maps. The encoder is a Hermitian vision transformer.

In practice

• Pretrain models on abundant unlabeled ultrasound IQ data.
• Employ Hermitian Vision Transformers for complex signal processing.
• Leverage self-supervision for label-scarce medical imaging tasks.

Topics

• IQ-JEPA
• Hermitian Vision Transformer
• Self-supervised Learning
• Ultrasound Imaging
• Sound Speed Estimation
• Label Efficiency

Best for: AI Scientist, Machine Learning Engineer, Research Scientist

Related on AIssential

• LatentWave: JEPA Pretraining for Wireless Foundation Models — JEPA pretraining for wireless foundation models enhances representation transferability by predicting masked regions in latent space.
• SHFormer: Dynamic Spectral Filtering Convolutional Neural Network and High-pass Kernel Generation Transformer for Adaptive MRI Reconstruction — SHFormer enhances MRI reconstruction by dynamically filtering spectral data and generating high-pass kernels for superior high-frequency detail.
• MIDI-RAE-JEPA: Hierarchical Representation Learning and Generation for Symbolic Music — MIDI-RAE-JEPA uses equivariance-based self-supervised learning to create rich, hierarchical representations for symbolic music generation and understanding.
• Contrastive Joint-Embedding Prediction for Representation Learning in Structural MRI — COJEPA combines JEPA and contrastive learning for robust, self-supervised 3D MRI representation learning.
• PolarBM: Complex-valued Boltzmann Machine for Modeling Audio Signals in Polar and Log-polar Coordinates — Modeling complex-valued data in polar coordinates, explicitly linking amplitude and phase, improves accuracy.
• STST-JEPA: Shallow-Target Spatio-Temporal Joint Embedding Prediction Architecture For EEG Self-Supervised Learning — STST-JEPA effectively predicts brain age and other traits from EEG using a novel self-supervised learning approach.

See Counsel's argued verdicts on the open AI decisions leaders are weighing →

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Machine Learning.