Photoacoustic (PA) imaging is a non-invasive medical imaging method that combines the high contrast of optical imaging with the deep tissue penetration of ultrasound. When a nanosecond pulsed laser illuminates biological tissue, the absorbed light energy causes local heating and rapid thermal expansion, generating broadband acoustic waves that are detected by ultrasound transducers and reconstructed into an image.
The magnitude of the initial acoustic pressure depends on the tissue's optical absorption coefficient, the local light fluence, and a tissue property called the Gruneisen parameter. Because different tissue types and molecules absorb light at different wavelengths, PA imaging can selectively visualize structures like blood vessels, tumors, and specific chromophores without external contrast agents.
Two primary PA imaging modalities have emerged. Photoacoustic microscopy (PAM) uses a focused transducer to scan point-by-point across tissue, producing high-resolution images of superficial structures like skin microvasculature. Photoacoustic computed tomography (PACT) uses a transducer array to capture signals simultaneously from many positions and then applies reconstruction algorithms to produce images of deeper structures including whole organs in small animals.
The choice between PAM and PACT involves a fundamental trade-off between resolution and depth. Optical-resolution PAM (OR-PAM) achieves sub-micrometer lateral resolution but is limited to shallow imaging depths. Acoustic-resolution PAM (AR-PAM) reaches tens-of-micrometers resolution at greater depth. PACT sacrifices some resolution but can image entire small animal bodies at spatial resolutions of hundreds of micrometers.
Deep learning refers to neural networks with many layers that automatically learn hierarchical feature representations from data during training, rather than relying on manually designed features. The core training loop involves feeding data through the network, comparing the output to ground truth using a loss function, computing gradients of the loss with respect to each network weight using backpropagation, and updating weights using gradient descent.
Convolutional neural networks (CNNs) dominate medical image analysis because their architecture reflects the structure of images. CNNs use convolutional layers with shared filter weights to extract spatially local features across the entire image, pooling layers that provide translational invariance, and nonlinear activation functions such as ReLU that allow the network to approximate complex nonlinear mappings.
Two regularization techniques are particularly important for medical imaging CNNs. Dropout randomly deactivates neurons during training, preventing overfitting by forcing the network to learn redundant representations. Batch normalization standardizes layer outputs within each training batch, stabilizing gradient flow and allowing much faster convergence without losing accuracy.
The three main categories of machine learning -- supervised learning (paired input-output training data), unsupervised learning (pattern discovery without labels), and reinforcement learning (learning through interaction and reward) -- are all represented in PA imaging research, with supervised learning currently dominant because most reconstruction problems can be formulated as mapping raw PA signals or noisy images to high-quality reference images.
Reconstructing a PA image from raw transducer signals is an ill-posed inverse problem: the standard universal back-projection (UBP) algorithm works well only when transducers cover sufficient angles and are densely sampled. In clinical scenarios, sparse transducer arrays and limited acquisition angles degrade image quality with streaking artifacts and missing structures. Deep learning addresses these problems more flexibly than traditional iterative methods.
Non-iterative reconstruction methods train a network to directly predict either the final image or an enhanced intermediate result in a single forward pass. Three distinct approaches have been developed: direct mapping from raw PA signals to images using architectures like U-Net and Res-UNet; enhancement of PA signals before reconstruction to correct for bandwidth limitations and noise; and post-processing of a conventionally reconstructed image to remove artifacts. The most common backbone architecture across all three approaches is U-Net, which combines encoder and decoder paths with skip connections.
Several specialized adaptations have been applied to PA reconstruction. DU-Net fuses PA signals acquired at three different transducer frequencies as separate input channels. Multi-wavelength approaches concatenate images from different laser wavelengths to enable spectral discrimination. Intravascular PA imaging uses networks to enhance cross-sectional vessel images generated from time-reversal reconstructions, with implications for arterial disease diagnosis.
For photoacoustic microscopy (PAM), deep learning corrects motion artifacts in OR-PAM, super-resolves under-sampled images acquired at reduced scanning density, and enhances images from low-cost LED-based PA imaging systems. LED systems produce much lower signal quality than pulsed laser systems due to limited power, but deep learning can recover image quality close to that of higher-power systems, potentially enabling affordable clinical deployment.
Iterative reconstruction methods embed deep learning inside an optimization loop that solves the PA inverse problem step by step. Rather than replacing the physics-based model entirely, these approaches use neural networks to learn better regularization terms or to correct the gradient steps in gradient descent, improving both convergence speed and final image quality compared to purely hand-designed algorithms.
A key example uses Recurrent Inference Machines to accelerate PA tomography reconstruction by learning the optimization trajectory from data. Another approach, the learned primal-dual algorithm, jointly reconstructs and segments vascular structures within a single architecture by learning from multi-task training supervision. These physics-informed methods tend to generalize better to unseen data than purely data-driven networks.
Beyond structure imaging, PA imaging can quantify blood oxygen saturation (sO2) by detecting hemoglobin's distinct absorption spectra at multiple wavelengths. Accurately measuring sO2 requires compensating for wavelength-dependent optical fluence variation in tissue, a confounding factor that simple linear unmixing cannot correct. Deep learning is well-suited to this problem: a ResU-Net trained on multi-wavelength simulated images demonstrated dramatically lower sO2 estimation error than linear unmixing, indicating the network learned to implicitly model the fluence correction.
Subsequent work on quantitative sO2 imaging has expanded from 2D to 3D PA images, incorporated joint blood vessel segmentation and quantification, and explored pixel-wise spectral input approaches rather than spatially global inputs. The field is moving toward unsupervised or self-supervised methods because ground truth sO2 measurements in living tissue are extremely difficult to obtain, making large labeled training datasets infeasible.
Deep learning for PA image detection aims to identify and localize pathological structures within PA images. Among the most significant clinical applications reviewed is prostate cancer detection using multispectral PA imaging. Arjun et al. applied a CNN combined with an adaptive feature selection strategy to classify ex-vivo human prostate tissue specimens, achieving the first reported use of deep learning for prostate cancer detection with PA imaging.
This initial 2D prostate cancer detection work was subsequently extended to 3D PA detection, using transfer learning from thyroid tissue datasets to compensate for the limited availability of prostate PA training data. A further study used multi-wavelength PA data to train a classifier distinguishing benign from malignant prostate tissue, demonstrating higher accuracy and sensitivity than classifiers trained on single-wavelength PA data alone.
A fully automated deep learning framework was also developed to detect and localize cancer regions in prostate tissue specimens from PA images without manual intervention. These prostate cancer detection studies collectively demonstrate a pipeline from tissue imaging to AI-driven diagnosis that could eventually reduce the need for invasive biopsies by enabling accurate in-vivo assessment of tissue malignancy.
PA image segmentation -- dividing PA images into anatomically meaningful regions -- has also progressed significantly. Sparse U-Net architectures automatically delineate blood vessels in multispectral optoacoustic tomography (MSOT), with integrated wavelength selection modules that reduce scanning time. Joint optoacoustic-ultrasound systems have used deep learning to segment animal boundaries for whole-body MSOT, and combined segmentation with quantitative functional imaging of hemoglobin oxygenation state in vessel walls.
PA-assisted intervention applies real-time PA imaging to surgical guidance, exploiting PA's ability to image both tissue structures and surgical instruments simultaneously without ionizing radiation. Deep learning enhances the image quality needed for this application by learning the physics of sound propagation to correct artifacts in PA images acquired during procedures.
One application targets neurosurgery, spinal fusion, and hysterectomy, where inadvertent damage to nearby blood vessels or nerves can cause paralysis or death. A deep learning system trained on PA images estimates the distance between surgical tools and critical structures in real time, providing the surgeon with a quantitative safety margin that is not visible under standard white-light imaging.
A second intervention application uses PA imaging for catheter tip tracking in cardiac procedures. The catheter tip generates a PA signal when illuminated by laser light, but reflections and other artifacts in the raw data obscure the tip's true location. A deep learning model trained to detect point sources in raw PA data accurately identifies the catheter tip position even when artifacts are present, prior to the beamforming step that produces the final image.
These surgical guidance applications benefit particularly from the real-time imaging speed of PACT systems and the ability of deep learning to process raw PA data faster than conventional reconstruction algorithms. The combination of high frame rates and AI-driven signal interpretation moves PA imaging closer to the operating room and interventional suite as a practical clinical tool.
The most fundamental challenge limiting progress in deep learning for PA imaging is the scarcity of standardized, publicly available datasets. Most PA imaging systems are still in prototype stages, and the variety of hardware configurations means that data collected on one system often cannot be used to train models for another. As a result, nearly all published studies validate on the same closed datasets used for training, preventing fair comparison between methods.
A related problem is the predominance of comparison against traditional iterative algorithms rather than other deep learning methods, making it difficult to determine genuine progress. The deep learning methods consistently outperform traditional baselines, but this comparison is increasingly uninformative. Standardized benchmarks analogous to the DIRLAB dataset in lung CT registration are needed to drive genuine competitive improvement.
Deep learning in PA imaging still suffers from the black-box problem -- it is unclear which features the networks learn and whether they are physically meaningful or represent statistical artifacts in the training data. For clinical translation, interpretability and uncertainty quantification are important: a clinician needs to know not only what the model predicts but how confident that prediction should be.
Looking ahead, the review anticipates that as PA imaging systems gain clinical approval and patient datasets accumulate, the field will shift from feasibility demonstrations to large-scale clinical validation. Integration of PA imaging with other modalities, particularly ultrasound, will expand available training data. The combination of physics-informed deep learning with PA's unique capability to measure functional parameters like blood oxygen saturation could eventually enable non-invasive metabolic tumor characterization without contrast agents.