光学技术在生物医学中扮演着越来越重要的角色，其非电离辐射、高分辨率、高对比度和对生物组织异变高度灵敏等特性使其非常适用于生物组织的研究，包括成像、传感、治疗、刺激以及控制等。然而由于光折射因子在生物组织中的分布是不均匀的，光在生物组织中的传播会受到很强的散射影响，故纯光学技术的穿透深度和空间分辨率是“鱼和熊掌不可兼得”；高分辨率光学成像应用仅限于样品浅表层，当成像深度增加时分辨率急剧下降。实现光在深层生物组织里的高分辨率成像或应用是人们期盼已久的目标。近年来，为解决这一问题，研究者提出了不同的方法，例如切换到更长的光波长以减小组织散射系数，在信号检测时将漫射光转换为散射不明显的超声信号，逆转或者预先补偿由光的多次散射所带来的相位畸变，或借助光纤等微创光学通道实现深层生物组织的高分辨率光学成像、刺激等。基于团队在深层生物组织光学相关领域多年的耕耘，从光在生物组织中的传播特性出发，梳理和总结了近年来研究人员在光-声结合和光学波前整形技术等方面展开的诸多探索，以及在生物组织操控、成像、光学计算以及人工智能等领域中的应用尝试。虽然尚有诸多不足，但随着硬件设备的更新和计算技术的发展，在不远的将来有望实现活体深层生物组织光学高分辨率应用。在这一求索过程中，新方法和新能力将不断激发新的应用灵感，为光学尤其是生物医学光子学带来全新的理念和机遇。

Optics, which is a significant sub-discipline of physics, focuses on the study of the phenomena, properties, and applications of light. Optics has evolved into an independent discipline over time. Optical imaging plays a crucial role in optical research by utilizing the phenomena and properties of light to record images of objects. Optical imaging has extensive applications in diverse fields, including astronomy, medicine, communication, and photography. For example, with the ongoing advancements in biomedical research, optical imaging has progressively showcased its distinctive advantages. First, optical imaging offers high resolution that is free from ionizing radiation, making it safer than X-rays or gamma rays that pose the potential risk of cancer. In addition, optical imaging can be flexibly configured to provide rich biomedical information based on the amplitude, phase, wavelength, polarization, and other characteristics of light. Another advantage of optics is their exceptional sensitivity, which enables the precise and sensitive detection of interactions between light and tissue components or molecules. Finally, the application of contrast agents further enhances the imaging specificity and contrast, thereby improving the visualization of desired targets and opening new avenues for disease diagnosis and treatment.

These have spurred the development of a vast range of high-resolution optical imaging technologies, such as confocal microscope, multiphoton microscope, and super-resolution imaging, which have been achieved by exciting fluorescence signals and/or utilizing gating or nonlinear optical effects in tissue samples. However, these implementations without exception have encountered fundamental challenges in thick biological tissues. This limitation stems from the strong scattering of light in tissue due to the inherent inhomogeneous spatial distribution of the refractive index of the medium encompassing diverse tissue constituents and functions. As a result, when light propagates within biological tissues, the light beam spreads quickly and is accompanied by the accumulated scattering of light (approximately one scattering event per 0.1-mm optical path length at visible wavelengths), which also rapidly weakens the intensity of non-scattered light in situ. In combination, these result in an intrinsic trade-off between spatial resolution and penetration depth for optics in biological tissues. This is also why optical techniques that utilize ballistic or quasi-ballistic photons typically have an effective penetration depth of less than or approximately 1 mm beneath the skin, which corresponds to 10 times the transport mean free path in the visible and near-infrared regimes. Excessive laser power may further enhance tissue penetration depths, but it also poses a risk of damaging biological tissues, particularly the skin and subsurface.

In the past two decades, numerous studies have been conducted to address these challenges, including switching to longer wavelengths to obtain lower tissue scattering coefficients, converting diffused light into non-scattered ultrasound at the signal detection side, and creating a minimally invasive optical path via ultrathin fibers to deep tissue regions. We believe that summarizing these advancements is not only worthwhile, but also critical for inspiring further research aimed at greater penetration depths and faster speeds toward wider applications.

In this review, we summarize the recent efforts in deep-tissue optics from various perspectives based on the mechanism of operation, including physical, computational, learning, and fiber optics. Note that this is not a complete list but only an empirical one.

Regarding physical-optics-based efforts, relevant research has primarily focused on the three aspects of wavelength engineering, energy conversion, and phase compensation. Wavelength engineering, such as multiphoton imaging and up-conversion imaging, involves the transformation of the input light wavelength into a different output wavelength to enhance the penetration depth. In multiphoton fluorescence imaging, two or more photons with longer wavelengths but lower energies are absorbed almost simultaneously before exciting the target fluorescent molecules at depth, generating one photon with shorter wavelength but higher energy. The longer wavelength in excitation and elevated photon energy in emission both contribute positively to the increased penetration depth for imaging. Up-conversion imaging entails the sequential absorption of multiple low-energy photons and their conversion into a single high-energy photon, thereby increasing the penetration depth.

Among approaches based on energy conversion, the photoacoustic (PA) effect, which converts input pulsed light into ultrasonic waves, has been extensively studied. When a biological tissue absorbs light energy, it undergoes thermal transformation, leading to localized expansion in the region of interest. Conversely, when the optical illumination is switched off, the local temperature decreases, causing the tissue region to contract. When the activation and deactivation of optical illumination (such as pulsed light) are manipulated, the expansion and contraction of tissues can be controlled, generating periodic mechanical waves in the ultrasonic frequency (MHz) range. These are usually referred to as photoacoustic or optoacoustic signals and are detected by one or an array of ultrasound transducers positioned outside the tissue sample. Because the generation of PA signals relies on the optical absorption of light, optical absorption contrast is obtained in PA imaging. However, the generation of signals does not distinguish between ballistic or diffused photons, and the detection of signals is based on ultrasound, which scatters much less (~1/1000) than light in the tissue. In combination, these features lead to a considerably boosted balance between imaging resolution and penetration depth and enable many exciting applications that are not possible with pure optical technologies.

In phase compensation, optical devices are utilized to measure and compensate for the optical phase distortion induced by light scattering. One representative example of phase compensation is optical phase conjugation, which captures the phase distortion of the wavefront emitted by a guide star within the scattering medium and compensates for it by conjugately adjusting the incident wavefronts and then refocusing light onto the position of the guide star. The phase-conjugation mirror, which is typically a photorefractive material, is responsible for recording the incident wavefront pattern and generating conjugated light that propagates along the optical path opposite the original transmission path.

Computational optics is an interdisciplinary field that merges optics and computers to leverage physics and algorithms, thereby enabling applications beyond those that can be achieved using traditional optical systems. The primary computational optics-based efforts in deep-tissue optics include digital optical phase conjugation (DOPC), iterative wavefront shaping, and transmission and reflection matrices. In DOPC, the phase-conjugation mirror previously discussed is replaced by the integration of a digital camera, computer, spatial light modulator, and algorithms for determining and generating the phase-conjugated wavefront. In iterative wavefront shaping, the phase of the incident light wavefront is adjusted based on feedback signals and the focusing performance is iteratively optimized. Feedback signals can take various forms, such as focal intensity, peak-to-background ratio (PBR) in the captured pattern, and photoacoustic signal strength. In the transmission matrix, a linear mathematical model is used to describe the relationship between the incident and scattered output wavefronts to characterize the scattering medium. If we denote the input wavefront as ein and the output wavefront as eout, the transmission matrix (MTM) can be characterized as eout=MTM⋅ein. By measuring the transmission matrix, we can focus the diffused light, project specific patterns through a scattering medium, or retrieve images from speckles. The reflection matrix establishes the relationship between the incident and reflected wavefronts from a scattering medium. In deep tissues, it is typically impractical to define or position guidestars or obtain guidestar signals within or on the opposite side of a tissue sample. Thus, applications of transmission matrices are limited. The introduction of a reflection matrix addresses this challenge by utilizing a reflected wavefront instead of a transmitted wavefront. In this scenario, both the incident and reflected light detectors are present on the same side of the scattering medium, thereby circumventing the need for guidestars to be placed on the opposite side of the scattering medium.

These computational optics-based efforts typically rely on intricate physical models to achieve the focusing or imaging of simple targets, such as letters, numbers, and other basic patterns, through scattering media. With recent advances in artificial intelligence, complicated problems involving speckles can now be addressed using deep learning. For example, deep-learning-based speckle imaging has powerful learning capabilities and data-driven characteristics. Deep neural networks can be trained using known data pairs, including ground-truth images and corresponding speckles, to extract various dimensions of information features. This can enable the high-fidelity reconstruction of target images, such as human face images. In addition, by training the speckle patterns obtained under different states of perturbed scattering media, the generalization capabilities of deep neural networks can be further improved, and the robustness of handling perturbed scattering media exceeds that of transmission-matrix-based methods.

In addition to these endeavors, which are all aimed at noninvasive deep-tissue optics, minimally invasive solutions that employ ultrathin optical multimode fibers as light guides into the tissue are also attractive and have seen promising advancements in recent years. Multimode fiber-based imaging is advantageous due to its minimally invasive nature, flexibility, and affordability. However, because of mode dispersion and coupling within multimode fibers, the optical field output from the fiber appears to be similar to a speckle pattern from tissue-like scattering media, making it infeasible to directly interpret the transmitted spatial information. Nevertheless, if multimode fibers are treated as scattering media, the aforementioned wavefront shaping approaches can be applied to multimode fibers. Thus, with the integration of wavefront shaping, the speckled output from a lensless multimode fiber can be focused onto a single optical mode, and then the raster can scan at a high speed within the field of view of the fiber. The excited or responding signals can also be detected and relayed using the same fiber for further use. This creates a scenario very similar to laser confocal microscope, except that the probe is inserted deep into the tissue. As a result, spatially and/or temporally resolving optical signals from deep tissues can be excited and detected with high resolution, which opens avenues for exciting new optical practices that require high resolution at depths in tissue. This capability can also be extended beyond imaging, such as for optogenetics, where wavefront shaping-empowered multimode fibers can deliver light precisely to targeted neurons within deep tissues and pick up fluorescence signals reflecting neuronal activities, enabling precise activation or inhibition of neurons to study brain functions.

Optics have gained significant attention in the study of deep biological tissues due to their non-ionizing radiation, exceptional contrast, exquisite specificity, and heightened sensitivity. In addition, the integration of computational optics and deep learning with conventional optics has substantially enhanced penetration depths while preserving moderate resolution in deep biological tissues. Despite these remarkable advancements, the practical implementation of deep-tissue optics still encounters critical challenges that must be addressed before moving forward.

The first is the penetration depth. With photoacoustic efforts and wavefront shaping techniques, which are sometimes further aided by computational optics and deep learning, current practices have achieved high-resolution optical focusing and/or imaging far beyond the optical diffraction limit. While most experimental research efforts to date still concentrate on small animal models such as mice, future studies are anticipated to improve the depth capability and extend to large animal models such as rabbits and monkeys. This transition is necessary for assessing the practicality, safety, and reliability of clinical diagnostics and therapeutic applications before working with human patients.

Speed is another crucial factor in the operation of deep-tissue optics. To reverse or compensate for the scattering-induced wavefront distortion, the scattering medium or multimode fiber should theoretically remain stationary to maintain the medium status, equivalent to the transmission matrix, during the wavefront optimization process. However, in practical applications, this requirement is hardly met, particularly for living biological tissues, whose optical field decorrelates rapidly on the order of milliseconds or even faster due to factors such as blood flow and respiration. Although some operations based on physical optics, such as optical phase conjugation, can reach this time scale, the majority of wavefront shaping implementations to date, consume seconds or hundreds of milliseconds, which is mainly limited by the response rate of the hardware such as spatial light modulators.

Over the past few years, deep learning has significantly affected deep-tissue optics. By leveraging the power of deep neural network models, it excels in extracting features and establishing nonlinear relationships between the target information (the ground truth) and the corresponding speckles, enabling high-fidelity retrieval of the original information from speckles. In addition, the use of deep learning has expanded the scope of speckle imaging, enabling breakthroughs in scattering, virtual staining, optical encryption, optogenetic networks, etc. The integration of deep learning with deep-tissue optics is expected to improve the speed, penetration depth, and immunity to system and medium disturbances. In addition, the combination of deep learning with physics-based scattering models holds great potential for accurately understanding and modeling multiple scattering processes, which is essential for designing efficient computation algorithms.

Finally, noninvasive deep-tissue optics in vivo still remains limited in some respects and may require a few more years to achieve technical maturity. Accordingly, a temporary yet effective alternative is to integrate wavefront shaping with ultrathin multimode fibers. Because the diameter of the multimode fiber can be 100‒200 μm, close to the typical hair diameter of adults, this integration can create a minimally invasive optical path into deep biological tissue, enabling high-resolution and fast-scanned optical focusing, imaging, stimulation, and manipulation at depths in tissue. Although it is not a perfect solution, it is practically useful in many studies, particularly for those at the early and preclinical stages, or when the insertion of a fiber-based probe is accompanied by invasive surgery, and the insertion of the probe does not considerably increase the degree of invasion or discomfort to the patient.

The developments to date in this field have demonstrated the feasibility and potential of deep-tissue optics. With continuing efforts and progress in related areas, technical barriers, such as the speed bottleneck associated with the response rate of spatial light modulators and the insufficient generalization capability of neural networks, can be overcome. It is strongly envisioned that in the near future, deep-tissue optics will reach practical maturity and be usable in vivo, which can extend many exciting optical applications to tissue regions that are currently optically inaccessible. This could reshape the landscape of light use in biomedicine and many other areas.