Deep Learning-Driven Quantitative Spectroscopic Photoacoustic Imaging for Segmentation and Oxygen Saturation Estimation

Spectroscopic photoacoustic (sPA) imaging can potentially estimate blood oxygenation saturation (sO2) in vivo noninvasively. However, quantitatively accurate results require accurate optical fluence estimates. Robust modeling in heterogeneous tissue, where light with different wavelengths can experience significantly different absorption and scattering, is difficult. In this work, we developed a deep neural network (Hybrid-Net) for sPA imaging to simultaneously estimate sO2 in blood vessels and segment those vessels from surrounding background tissue. sO2 error was minimized only in blood vessels segmented in Hybrid-Net, resulting in more accurate predictions. Hybrid-Net was first trained on simulated sPA data (at 700 nm and 850 nm) representing initial pressure distributions from three-dimensional Monte Carlo simulations of light transport in breast tissue. Then, for experimental verification, the network was retrained on experimental sPA data (at 700 nm and 850 nm) acquired from simple tissue mimicking phantoms with an embedded blood pool. Quantitative measures were used to evaluate Hybrid-Net performance with an averaged segmentation accuracy of >= 0.978 in simulations with varying noise levels (0dB-35dB) and 0.998 in the experiment, and an averaged sO2 mean squared error of <= 0.048 in simulations with varying noise levels (0dB-35dB) and 0.003 in the experiment. Overall, these results show that Hybrid-Net can provide accurate blood oxygenation without estimating the optical fluence, and this study could lead to improvements in in-vivo sO2 estimation.

💡 Research Summary

Spectroscopic photoacoustic (sPA) imaging offers a non‑invasive route to estimate blood oxygen saturation (sO₂) by exploiting the wavelength‑dependent optical absorption of hemoglobin. Accurate quantification, however, traditionally requires solving the optical fluence inverse problem, which is challenging in heterogeneous tissues where scattering and absorption vary strongly with wavelength and depth. In this work, the authors introduce Hybrid‑Net, a deep neural network that simultaneously segments blood vessels and predicts their sO₂ directly from dual‑wavelength sPA images, thereby bypassing explicit fluence estimation.

Hybrid‑Net builds on a U‑Net encoder‑decoder architecture with skip connections, accepting a 128 × 128 × 2 input (700 nm and 850 nm sPA images). The network outputs two 128 × 128 maps: a binary vessel segmentation (1 = vessel, 0 = background) and an intermediate sO₂ map defined over the whole field of view. The final sO₂ estimate is obtained by element‑wise multiplication of the segmentation and intermediate sO₂ maps, ensuring that only voxels identified as vessels contribute to the oxygenation measurement. The loss function combines a Dice + binary‑cross‑entropy term for segmentation with a mean‑squared‑error term for sO₂, applied only within the segmented vessel region—a “hybrid loss” that forces the network to focus sO₂ learning where it matters.

Training proceeds in two stages. First, a large synthetic dataset is generated using three‑dimensional Monte‑Carlo light‑transport simulations of a 38 mm³ breast‑tissue volume containing epidermis, dermis, and fat layers, plus randomly placed cylindrical blood vessels (radius 0.5–4 mm) with random orientations and sO₂ values. For each vessel configuration, 700 nm and 850 nm optical excitations are simulated with 10⁸ photons, producing initial pressure distributions that are then reconstructed into 2‑D sPA images. Gaussian noise is added to achieve signal‑to‑noise ratios (SNR) ranging from 0 dB to 35 dB, yielding 4 000 paired images (2 000 per wavelength). Hybrid‑Net is pre‑trained on this dataset using Adam (learning rate = 1e‑4) for 150 epochs with early stopping, batch normalization, and 0.1 dropout to improve convergence and prevent over‑fitting.

Second, the network is fine‑tuned on experimental data acquired with a custom ultrasound‑photoacoustic system. Tissue‑mimicking phantoms consist of a gelatin matrix (Hummic Medical) containing an embedded blood‑flow channel. Blood oxygenation is modulated by bubbling CO₂ and O₂ gases, creating a range of sO₂ values. Dual‑wavelength (700 nm/850 nm) PA signals are recorded using a 128‑element linear US transducer synchronized with a 4 cm diameter optical fiber illumination. A total of 410 PA image pairs (each representing a distinct vessel location and sO₂ level) are collected; 80 % are used for re‑training, 10 % for validation, and 10 % for testing. Data augmentation (random rotations, flips, and spatial shifts) expands the training set to 1 640 samples.

Performance is evaluated both on simulated data with varying noise levels and on the experimental phantom. In simulation, Hybrid‑Net achieves an average Dice coefficient of ≥ 0.978 for vessel segmentation and an sO₂ mean‑squared‑error (MSE) of ≤ 0.048 across all SNR conditions. In the phantom experiments, segmentation accuracy reaches 0.998 and sO₂ MSE drops to 0.003, indicating near‑perfect recovery of oxygen saturation without any explicit fluence correction. These results surpass conventional linear unmixing or model‑based approaches that rely on fluence estimation, demonstrating that a data‑driven network can implicitly learn the complex wavelength‑dependent light transport in heterogeneous media.

The paper’s contributions are threefold: (1) a unified deep‑learning framework that jointly performs vessel segmentation and quantitative sO₂ estimation from only two wavelengths; (2) a hybrid loss that isolates sO₂ error to the segmented vessels, improving robustness against background noise; and (3) a two‑stage training pipeline that leverages large‑scale Monte‑Carlo simulations for pre‑training and experimental phantom data for domain adaptation.

Limitations include the reliance on only two wavelengths, which may restrict spectral discrimination in more complex chromophore mixtures, and the fact that the network has been validated only on simple phantoms rather than in vivo tissue. Additionally, very thin or highly tortuous vessels could challenge the segmentation accuracy. Future work should explore multi‑wavelength extensions, three‑dimensional network architectures, and validation on animal or clinical datasets to assess real‑world applicability.