Hybrid photoacoustic and fast super-resolution ultrasound imaging

Dual-modality PA/US imaging leverages complementary contrasts: PA provides high-contrast vascular imaging without exogenous agents, hemoglobin oxygenation, lipid/collagen content, and molecular contrast; US offers blood perfusion, morphology, and elasticity, with shared hardware enabling interleaved acquisition. Despite versatility, diffraction-limited resolution and penetration trade-offs constrain deep-tissue resolution. Super-resolution localization techniques (ULM and PA localization) can break the acoustic diffraction limit, but PA localization is limited by weaker signals and special agents. Conventional ULM requires low microbubble densities to avoid PSF overlap and demands thousands of frames, resulting in tens to hundreds of seconds per frame even with ultrafast US, hindering real-time dual PA/ULM. The study aims to accelerate ULM using sparsity-constrained optimization to enable fast interleaved PA/ULM imaging with improved temporal resolution, reduced motion artifacts, and robust 3D/time-lapse dual-contrast imaging in vivo.

Prior dual PA/US studies demonstrate broad preclinical and emerging clinical utility (breast, melanoma, sentinel lymph node, endoscopy, brain). ULM provides up to a ten-fold theoretical resolution improvement over conventional US. Standard localization (e.g., PSF cross-correlation, PSF-CC) requires low-density microbubbles to avoid overlap, leading to long data acquisition. Compressed sensing approaches can localize overlapped microbubbles and speed ULM but often rely on computationally heavy patch-stitch processing, potentially affecting accuracy. Deep-learning localization requires large organ-specific training datasets. Sparsity-based super-resolution correlation imaging avoids explicit localization but had not delivered super-resolved velocity maps. The need remains for a fast, accurate, computationally efficient ULM method compatible with interleaved PA imaging and robust to higher bubble densities.

System and acquisition: A dual-modal PA/US system used a Verasonics Vantage 256 with a 15 MHz linear array (VisualSonics MS200/MS250) and a wavelength-tunable 690–950 nm OPO laser (7 ns pulses, 10 Hz). A custom bifurcated fiber bundle provided side illumination. Interleaved imaging cycles alternated multi-wavelength PA (0.1–0.5 s per cycle at 10 Hz; each wavelength 0.1 s) and ultrafast plane-wave US for ULM (0.5–1.5 s), enabling <2 s per position in 3D scans. Plane-wave compounding used seven angles (−6° to 6°) at 500 Hz PRF.

SC-ULM with FISTA: After high-pass spatiotemporal clutter filtering and thresholding on IQ data to isolate microbubble signals, each ultrasound image b was modeled as b = DHx + w, with D a downsampling operator (factor 8), H a PSF-induced blurring operator (BCCB matrix), x the sparse microbubble distribution, and w noise. Microbubble locations were recovered via nonnegative ℓ1-regularized least squares: minimize ||Ax − b||² + λ||x||₁ subject to x ≥ 0, with A = DH. A Gaussian PSF (spatially invariant approximation) was estimated from isolated bubbles. The optimization was solved using FISTA with a nonnegative shrinkage operator. Efficient application of Aᵀ and AᵀA was achieved via FFTs exploiting H’s BCCB structure, avoiding explicit storage of large matrices and patch-stitching. For comparison, CVX and L1-homotopy sparsity solvers and standard PSF-CC localization (normalized 2D cross-correlation with coefficient threshold 0.6) were implemented.

Tracking and velocity estimation: After localization, rigid-motion correction (see below) was applied. Microbubbles were tracked frame-to-frame using a Hungarian-method-based particle tracker (simpletracker.m), forming trajectories to compute velocities and construct super-resolved images and speed maps.

Simulations: Using the Verasonics Ultrasound Simulator, microbubbles (Gaussian amplitude distribution) were randomly placed in a 2.5 × 2.5 mm field. A 15 MHz linear array transmitted seven-angle plane waves at 500 Hz; RF data were reconstructed (delay-and-sum) into 1000 B-mode frames. Experiments covered various microbubble densities (N=20 repeats each). Localization performance (identification rate, localization error) was compared between SC and PSF-CC.

In vivo studies: Six female BALB/cJ mice (8–10 weeks) were used under IACUC approval. For hindlimb lymph node imaging, 20 µL ICG (0.2 mg/mL) was injected in the footpad; after 15 min, 200 µL lipid microbubbles (1×10⁸ particles/mL; mean size ~1.03 µm) were injected IV. A 3D scan acquired 16 lateral positions (0.1 mm step). At each position: ultrafast US 1.5 s (500 Hz PRF, seven angles) for ULM; PA 0.3 s at 750, 790, 850 nm. For kidney imaging (left kidney), mice were anesthetized and injected IV with 100 µL microbubbles (1×10⁸/mL). Interleaved ultrafast US (0.8 s) and PA (0.2 s) acquisitions were repeated during an oxygen-challenge protocol: 100% O₂ for 50 s, then 3% O₂/97% N₂ for 50 s, then 20% O₂/80% N₂ for 50 s. US central frequency was 15 MHz, frame rate 500 Hz.

Motion compensation: Respiratory frames were identified by frame-to-frame normalized cross-correlation on the B-mode stack; frames below a 0.99 threshold were discarded as out-of-plane respiration motion. Remaining frames were grouped by respiratory cycles. Inter- and intra-cycle in-plane motions (lateral/axial translations and rotation) were corrected using rigid registration (MATLAB imregtform with regular step gradient descent). Residual inter-cycle out-of-plane motions were not correctable; minimizing DAQ duration per frame reduced their impact.

PA spectral unmixing: After fluence normalization (from measured pulse energy and illuminated area), linear spectral unmixing separated chromophore contributions (HbO2, Hb; plus ICG for hindlimb) using known molar extinction spectra, yielding oxygen saturation maps (sO2 = HbO2/(Hb+HbO2)).

Image analysis: Saturation curves quantified the fraction of explored pixels in vascular ROIs versus acquisition time; characteristic time was defined by the initial slope extrapolated to full saturation. Resolution was measured by Fourier ring correlation (half-bit threshold) using split-track images. Statistical analyses used unpaired two-tailed t-tests (p<0.05 significant).

• Synthetic data: At 16 microbubbles/mm² (100 bubbles in 6.25 mm²), SC recovered 92% of bubbles vs 35% for PSF-CC. As density increased (e.g., 85, 142, 192 bubbles/mm²), SC maintained higher identification with a gradual linear decrease (>80% under 100/mm²; >60% under 200/mm²), whereas PSF-CC fell to <5% above 75/mm² and <2% at 200/mm². SC recovered up to 37× more bubbles than PSF-CC at high densities and achieved lower localization error across densities, though both degraded with increasing density.
• Computational efficiency: The SC-FISTA approach was up to 57× faster than CVX and up to 10× faster than L1-homotopy on test sizes (25×25 input to 200×200 output); with larger images, SC could be up to 137× faster than CVX. SC also yielded greater localization accuracy than CVX (6×) and L1-homotopy (5×) in supplementary benchmarks.
• In vivo hindlimb: SC-ULM produced brighter, more complete microvascular maps than PSF-CC for the same DAQ. Saturation analysis: characteristic time 1.5 s for SC-ULM vs 32.5 s for PSF-CC (≈22× faster). At 1.5 s DAQ, FRC resolutions were 36.6 µm (SC-ULM) and 35.8 µm (PSF-CC); resolution differences became insignificant after 1.5 s. SC revealed major vessels by 0.5 s; PSF-CC remained sparse at 7 s.
• 3D dual imaging of lymph node: Interleaved 3D acquisition with <2 s/position (1.5 s ULM + 0.3 s PA) over 16 positions enabled co-registered maps of ICG-labeled popliteal lymph nodes and nearby microvasculature, including lymph node microvasculature.
• In vivo kidney: SC-ULM reconstructed renal microvasculature within 1 s; PSF-CC remained sparse at 1 s and blurred by 4 s due to physiological motion. Kidney saturation characteristic time: SC ≈0.5 s vs PSF-CC >13 s (≈28× faster). Motion correction: 1 s SC images showed minimal improvement from in-plane correction (MS-SSIM 0.97±0.005), whereas 4 s SC images required in-plane correction (MS-SSIM 0.84±0.024) yet remained affected by out-of-plane motion. SC reduced motion-correction complexity by fitting within single respiratory cycles.
• Physiological mapping: Interleaved imaging during oxygen challenge (100%, 3%, 20% O₂) produced co-registered renal hemoglobin oxygenation (PA) and blood speed maps (SC-ULM), revealing concurrent changes and a positive correlation between blood flow speed and hemoglobin oxygenation in selected regions.

The study addresses the key barrier to dual PA/ULM imaging—ULM’s long acquisition time—by introducing a sparsity-constrained FISTA-based localization that tolerates higher microbubble densities and recovers many more events per frame. This dramatically reduces DAQ per super-resolved frame (to ≤1.5 s, as low as 0.5–1 s in kidney), enabling interleaved acquisitions within a single respiratory cycle. Consequently, untrackable out-of-plane motions are minimized, simplifying motion correction and improving spatial-temporal co-registration of PA and ULM. Synthetic and in vivo results show that SC-ULM maintains high identification rates and acceptable localization precision at densities where PSF-CC fails, yielding up to 22–28× speedups in vivo and enabling practical 3D dual imaging with a standard linear array system. Demonstrations in lymph node imaging highlight co-localization of dye-labeled lymphatic structures with surrounding microvasculature, and renal studies show simultaneous mapping of hemodynamics and oxygenation with consistent co-registered dynamics across oxygenation states. These findings substantiate the method’s potential to expand dual-modality imaging to applications requiring fast, super-resolved vascular information and physiological readouts.

The work presents an interleaved dual PA/super-resolution US imaging technique accelerated by sparsity-constrained optimization (FISTA), achieving super-resolved vascular imaging at sub-2 s per frame and enabling 3D and time-lapse dual-contrast studies with reduced motion-correction burden. Key contributions include: fast SC-ULM robust to higher bubble densities (up to 37× more recoveries in simulation) with 22–28× in vivo speed gains over PSF-CC; practical 3D co-registered imaging of ICG-labeled lymph nodes and microvasculature; and simultaneous renal blood flow and oxygenation mapping during oxygen challenges. Future directions include increasing PA frame rates via higher-repetition lasers or multi-laser setups, accelerating clutter filtering (e.g., randomized SVD and spatial downsampling), and GPU-accelerated FISTA for real-time performance, potentially broadening impact in cancer, neuroscience, nephrology, and immunology.

• Trade-off between temporal and spatial resolution: Higher bubble densities and shorter DAQ improve speed but reduce localization precision for both SC and PSF-CC.
• PSF assumptions: A spatially invariant Gaussian PSF was used; real PSFs vary spatially and with bubble proximity, introducing recovery errors.
• Tissue noise and false positives: In vivo, tissue signals can be misidentified as bubbles; SC’s greater sensitivity can increase localization errors early in acquisition, slightly reducing resolution relative to PSF-CC at very short DAQs.
• Motion artifacts: Although interleaved fast acquisition mitigates out-of-plane motion, longer acquisitions (e.g., multi-cycle) still suffer from uncorrectable inter-cycle out-of-plane shifts, especially in kidney imaging.
• Nonuniform saturation: Large vessels saturate earlier than small ones, biasing early saturation metrics and transiently lowering apparent resolution until sufficient DAQ accrues.
• Hardware constraints: PA frame rate was limited by a 10 Hz laser; higher-repetition-rate sources are needed for real-time dual imaging.