IQ-JEPA — A Joint-Embedding Predictive Architecture for Quantitative Ultrasound
— by Vivax
IQ-JEPA brings Yann LeCun's Joint-Embedding Predictive Architecture to medical ultrasound.
On July 24, 2026, Masashi Sode and Gianmarco Pinton published IQ-JEPA on arXiv — the first Joint-Embedding Predictive Architecture built for raw ultrasound channel data. The target is quantitative ultrasound's hardest everyday problem: estimating the speed of sound and attenuation inside tissue. Sound speed is a prerequisite for well-focused imaging and carries diagnostic value in its own right, but recovering it from pulse-echo data is a nonlinear inverse problem. The field's learned solvers are fast but label-hungry — and the labels come from expensive physics simulations, while hospitals sit on abundant unlabeled channel data. IQ-JEPA is designed to exploit exactly that asymmetry.
The recipe is pure JEPA, transplanted from natural images to physics. During pretraining, no labels are used at all: regions of the in-phase and quadrature (IQ) signal are masked, and an encoder must predict the latent representation of the hidden regions from the visible context — never the raw signal itself. Only afterwards is the model fine-tuned on simulated sound-speed maps. This is the same masked latent-prediction objective that powers Meta's I-JEPA and V-JEPA, applied to a signal no consumer camera ever sees.
The architectural contribution is a Hermitian vision transformer that operates directly on the complex IQ signal. Sound speed shows up in IQ data as a phase difference that is invariant to any constant phase offset — so the encoder is built to respect that physics: its attention mechanism is equivariant to the phase, and its conjugate-product feed-forward layers are invariant to it. The result is a network that reads the same quantity classical coherence-based methods use, but learns it end-to-end instead of hand-deriving it.
The evaluation runs on 79,293 Fullwave 2.5 simulations at 2.5 MHz, with 63,435 acquisitions reserved as unlabeled pretraining data. The headline is label efficiency: with JEPA pretraining, the model reaches 15.60 m/s sound-speed error using 10,000 labels — roughly a threefold gain in label efficiency over purely supervised training, and the gain grows past fourfold when only 1,000 labels are available. With full labels the error drops to 8.71 m/s, and the model sits about 2.2× ahead of an InversionNet baseline. Two further results matter as much as the headline: the label-efficiency gain keeps growing with more unlabeled pretraining data, and the authors' ablations point to self-supervision — not the architecture alone — as the dominant factor.
The transfer results are why the authors frame this as a first step toward a foundation model for quantitative ultrasound. The pretrained encoder's frozen features already expose both sound speed and attenuation without task-specific training, and pretraining on layered phantoms transfers to abdominal phantoms (and vice versa) at little cost in accuracy. That is the foundation-model signature: one self-supervised backbone, many quantitative readouts. IQ-JEPA joins a fast-growing medical JEPA lineage — alongside Neuro-JEPA for multimodal brain MRI — extending Yann LeCun's latent-prediction blueprint from internet video into clinical physics.
IQ-JEPA is the clinical JEPA thesis in miniature, proven on real physics. Medicine's data economics are exactly the ones this paper exploits: unlabeled signal is abundant, expert or simulated labels are scarce and expensive, and the quantity that matters is hidden beneath the raw data's surface. A threefold-to-fourfold label-efficiency gain is not an incremental benchmark point — it is the difference between a model a hospital can afford to specialize and one it cannot. This is precisely how Vivax builds its clinical world models: pretrain on the data that already exists, respect the domain's physics in the architecture, and spend the scarce labels only where they buy the most. Every new domain JEPA conquers — images, video, robots, brain MRI, now raw ultrasound — is another data point that the latent-predictive route is the road medical AI will travel.