Achieving a cellular-resolution understanding of organ or whole-body architecture is a fundamental objective in biology. Recent advancements in tissue clearing techniques have enabled unprecedented visualization of entire or partial body structures, e.g., via fluorescent imaging, highlighting the importance of this approach^[1–4]. These techniques enhance tissue transparency by minimizing light scattering, which arises from material inhomogeneities with varying refractive indices (RIs) and the spatial distribution of scattering elements^[5,6]. Tissue clearing achieves this by employing methods such as removing or diluting scattering-inducing components (e.g., membrane lipids), altering tissue morphology through swelling or shrinkage to modify the extracellular matrix (ECM), or matching the RI of the medium to the average RI of tissue components. Existing protocols are designed around these principles. Since organs and tissues exhibit distinct macromolecular compositions and histological characteristics, their optical properties and responses to clearing methods vary.

To date, optical clearing methods have been applied to nearly all animal tissues and numerous plant specimens^[1,7,8]. Despite the significant interest and potential of these techniques, the field remains largely qualitative rather than quantitative in terms of method selection. Although comparative studies have been conducted for nearly every organ to identify the most effective clearing technique^[9–16], these assessments often rely on unstandardized and sometimes subjective criteria.

Only a few research groups have proposed systematic, reproducible approaches to evaluating the effects of tissue clearing across different methods. For instance, Woo et al.^[17] developed the Punching-Assisted Clarity Analysis (PACA)-Light and PACA-Glow methods, which utilize widely accessible spectroscopy and gel documentation systems to quantify tissue transparency. This approach enabled quantitative comparisons of 28 tissue-clearing protocols in rodent brains. However, it requires sampling before and after a predefined (and subjective) incubation period without the capability to monitor the process in real time. Similarly, Kim et al.^[18] designed assay systems to analyze clearing properties across tissues and evaluate tissue-clearing quality based on changes in macromolecular composition and transparency. Their system facilitates both the selection of an optimal method for a specific tissue type and subsequent quality control to ensure reproducibility. However, it is based on an advanced electrophoresis-driven clearing system, which limits its accessibility for laboratories without prior expertise in tissue clearing techniques. Additionally, it is very hard to achieve high spatial and temporal resolution of an optical clearing assay.

In this paper, we present a new, efficient, and user-friendly method, called tissue clearing AI monitoring (ClearAIM) for real-time high space-bandwidth product monitoring of the optical clearing process leveraging deep learning based dynamic segmentation scheme (segment anything model, SAM^[19]). We demonstrate its application using a cost-effective easy-to-build setup (all off-the-shelf components) and mouse brain slices cleared with the widely utilized CUBIC method. The proposed system is designed to provide users with (1) precise data on the level of transparency achieved and the time required to reach it (toward saturation of the process), (2) real-time insights into morphological changes during the clearing process, and (3) an optimized timeline for clearing that balances adequate tissue transparency for fluorescent imaging (e.g., confocal microscopy) with minimal morphological alteration caused by prolonged incubation. This approach seeks to enable the informed selection and accurate tuning of tissue clearing methods to be a practical, efficient, and reproducible process across a wide range of laboratory settings.

The proposed ClearAIM method consists of two major steps: (1) time-lapse imaging in a brightfield microscopic setup with demagnification, described in Sec. 2.1, and (2) automatized digital analysis of all collected intensity images for quantitative determination of the tissue area and transparency over time, described in Sec. 2.2. Samples were prepared according to the description provided in Sec. 2.3.

The optical setup used for the measurements is illustrated in Fig. 1(a). The sample was illuminated by a white light LED positioned approximately 25 cm below the sample. Lens L1 ($f=150\mathrm{mm}$) was placed 9 cm above the sample, while lens L2 ($f=80\mathrm{mm}$) was positioned 40 cm above L1. Together, these lenses formed an imaging system that projected the sample’s image onto a camera (DAHENG IMAGING MER2-230-168U3M, resolution $1920\text{pixel}\times 1200\text{pixel}$, pixel size 5.86 µm) mounted 8 cm above L2. The system was designed to operate at a magnification of $0.35\times$, providing a field of view (FOV) of $32\mathrm{mm}\times 20\mathrm{mm}$, sufficient to fit an entire mouse brain tissue slice. The resolution of 15.63 µm, measured experimentally using the USAF target, was considered adequate (and much higher than that in Ref. [17]) for the primary objective of tracking global changes in the sample’s transparency during the optical clearing process.

The main goal of the setup was to image the sample plane and monitor the optical clearing process over time. To maintain a uniform background, a Petri dish with a glass bottom (IBIDI µ-dish 35 mm, high glass bottom) was used. This dish featured a favorable design, comprising a plastic wall and a glass coverslip adhered to the bottom. The glass bottom was slightly recessed relative to the surrounding plastic ring. This height difference raised concerns that the tissue sample might roll during the clearing process, potentially introducing artifacts or false observations. To address this issue, a custom-designed spacer was fabricated using 3D printing. The spacer was designed with an internal diameter (20 mm) matching the glass bottom, an external diameter fitting the inner walls of the Petri dish, and a height of 5 mm. Additionally, the spacer confined the setup’s effective FOV to its internal dimensions.

The sample is prepared for measurement as illustrated in Fig. 1(b). First, the spacer was placed inside the Petri dish, followed by the tissue sample, which was then submerged in the CUBIC-R1 clearing solution, ensuring the spacer was fully immersed. To minimize evaporation, the upper part of the Petri dish was secured. Data acquisition commenced immediately after the sample was placed in the Petri dish, as the optical clearing reaction begins upon the addition of CUBIC-R1. The frequency of data collection was adjusted based on the sample’s thickness, given that thinner samples (e.g., 500 µm) exhibit faster clearing dynamics than thicker ones (e.g., 1 mm). For a 500-μm-thick sample, images were captured every second during the first 5 min, followed by one image per minute for several hours, and subsequently every 10 min until the experiment concluded. For a 1-mm-thick sample, images were taken every second for the first 5 min, every minute for the next hour, and then every 30 min for the remainder of the experiment.

To dynamically analyze the evolving morphology of biological tissue undergoing the optical clearing process, a time-varying mask is needed to enclose the object (tissue slice) and allow for calculation of its area and transparency. To address this challenge, we developed a segmentation algorithm combining a robust deep learning model with an original adaptive frame-to-frame guidance scheme. This approach ensures consistent segmentation across sequential frames and minimizes errors arising from changes in sample transparency, morphology, or external factors such as variable illumination conditions. The proposed ClearAIM is freely available on the GitHub repository^[20].

The workflow, as shown in Fig. 2, begins with the manual initialization of the region of interest (ROI) in the first frame (consisting of the object and surrounding background). The mask generated from this initial step serves as a temporal guide for subsequent frames, as shown in Fig. 2(a). The segmentation (logical mask generation) process employs the Segment Anything Model (SAM, ViT-H version)^[19], pre-trained on diverse datasets, as the backbone, ensuring robustness and adaptability to a wide range of sample conditions. The SAM produces high-quality object masks from input prompts such as points (inside objects to be segmented), and it can be used to generate masks for all objects in an image. Image preprocessing is implemented to ensure the reliability of all the following steps. Input images are first downsampled to reduce computational overhead while retaining essential structural information. Pre-scaling parameters are to be set empirically to balance computational efficiency and segmentation accuracy.

The key step of ClearAIM is the automatic generation of a logical mask indicating the region of a tissue sample for each captured frame. This is done sequentially with SAM based on the current frame and prompts positive and negative points that help to differentiate between the sample and background, respectively. For the first frame, positive and negative prompt points are initialized manually. We have, thus, two important parameters: (1) a number of positive points to identify the object of interest and (2) a number of negative points to define the object’s boundaries. Recommended—highly sample dependent—values, tested empirically, are 2 points for the sample region and 20 points for the background (sample-free). For subsequent frames, positive points are derived from stable regions of the overlap between the current and previous masks, while negative points are determined by inverting the current mask to exclude irrelevant areas. Instead of manual point selection, we use a clustering technique (K-means) to analyze the previous image’s mask to adaptively place the new points. Masks are predicted with SAM for each frame based on the current image and prompt inputs. The algorithm identifies the largest connected region in each mask as the tissue sample and in this way excludes artifacts, such as debris or reflections, from further analysis.

These masks form the basis for downstream quantitative analyses, such as the sample area (calculated as the sum of pixels in the mask area) and transparency (calculated as the ratio of the mean intensities in the object and background areas), as shown in Fig. 2(b). To minimize the impact of early experimental disturbances, transient surface effects, and intensity fluctuations (e.g., due to the CUBIC-based lipid extraction from the tissue slice), we used the local background mean intensity as a reference for transparency calculation for each frame instead of the initial time point ($t=0$). Morphological operations are used for processing masks, including detecting the object’s surroundings and preprocessing (e.g., filling small gaps in masks or removing edge artifacts). All logical mask detection operations are performed in Python, while all tasks related to analyzing the original image and mask to obtain measurements and create animations are conducted in MATLAB.

The use of the previous frame’s mask as a guidance prompt introduces temporal consistency, reducing segmentation drift and improving robustness. This approach ensures reliable tracking of the sample’s morphology over extended time-lapse sequences, even under significant fluctuations in brightness, contrast, or transparency. The method was specifically designed to address the challenges posed by dynamic sample conditions during tissue-clearing procedures, such as those involving CUBIC-treated samples. By integrating temporal information, robust prompt generation, and efficient processing, the algorithm provides high-precision segmentation results with minimal user intervention, enabling researchers to extract meaningful metrics over time.

Tissues were obtained from inbred C57BL/6J mice (bred at the Center for Experimental Medicine in Białystok and the M. Mossakowski Medical Research Center in Warsaw). The animals were provided a controlled environment (temperature 24°C, 12/12 light/dark cycle) with ad libitum access to water and feed. The animals were euthanized by cervical dislocation following exposure to isoflurane (FDG9623, Baxter). Isolated tissues were fixed for 24 h in a 4% paraformaldehyde solution (BD Cytofix/Cytoperm™, No. 554722, BD Biosciences) at 4°C before being used in the experiments. All procedures were conducted in accordance with the Directive of the European Parliament and Council (2010/63/EU) and Polish Law. Mice were solely tissue donors, meaning no ethical committee approval was required.

The fixed brains were embedded in a 3% w/v agarose solution (Sigma-Aldrich) in water and subsequently sectioned to the desired thickness using a Vibratome (VT1000S, Leica). The sections were stored in $1\times$ phosphate-buffered saline (PBS, Sigma-Aldrich) supplemented with 0.05% sodium azide (NaN3, Sigma–Aldrich).

The CUBIC-R1 solution was prepared according to a published protocol^[21]. In brief, a solution was prepared as a mixture of 25% (mass fraction) urea (Sigma-Aldrich), 25% (mass fraction) $N,N,N^{\prime },N^{\prime }$-tetrakis(2-hydroxypropyl)ethylenediamine (Sigma-Aldrich), and 15% (mass fraction) Triton X-100 (Sigma-Aldrich) dissolved in a distilled water. For tissue optical clearing, the agarose was removed from the brain slices, which were then submerged in an excess (around 5 ml) of CUBIC-R1 solution in a custom-made Petri dish to minimize the potential effects of solution evaporation.

The 500 µm brain slices were stained with propidium iodide (at a concentration of 7.5 µg/ml dissolved in PBS) for 24 h prior to CUBIC-R1 clearing. For imaging, a Leica TCS SP8 confocal microscope (Leica Microsystems) equipped with an HCX PL APO CS $10\times /0.40$ objective was employed. Excitation was achieved using a white light laser at an appropriate wavelength for propidium iodide. The acquisition was performed using a hybrid detector (HyD) at a resolution of $1024\times 1024$, with a speed of 400 Hz and a $z$-step of 2 µm under the same laser power for all conditions. For calculating slice-per-slice image background fluorescence, imaging data were analyzed similarly to Nürnberg et al.^[22] in ImageJ. Due to the very low signal in the lower $\sim 100\mathrm{\mu m}$ of the imaged stacks, the analysis considers values up to 400 µm.

Figure 3 presents the evaluation results of the proposed ClearAIM method for monitoring the optical clearing process using the CUBIC-R1 solution on two exemplifying mouse brain tissue slices of different thicknesses: 500 µm and 1 mm. Both tissue slices were imaged using the described setup over a total duration of 71 h. Figures 3(a1), 3(a2) and 3(b1), 3(b2) show the slices at the start and end of the clearing process, for 500-μm- and 1-mm-thick slices, respectively. Additionally, Visualization 1 and Visualization 2 provide time-lapse sequences of the acquired images, capturing the clearing process for the 500-μm- and 1-mm-thick slices, respectively. From the videos and images, it is evident that both tissue slices expanded in size and became more transparent during the clearing process. This observation is further supported by quantitative analysis using the proposed ClearAIM (Fig. 2) method, which enabled the measurement of tissue transparency [Fig. 3(c)] and the cross-sectional area [Fig. 3(d)] over time. In Fig. 3(e), data from Figs. 3(c) and 3(d) are combined to provide an integrated visualization of the interplay between tissue slice transparency and size variation. To facilitate direct comparison, the measured values within each video’s timeframe were normalized to a [0–1] range using the minimum and maximum values obtained from the respective datasets. Furthermore, Fig. 3(d) was normalized based on the maximum time-lapsed size, thereby offering a representation of relative sample expansion that is largely independent of initial measurement fluctuations.

Figure 3 shows that the transparency of the 500 µm slice increased from 43% to 62%, while the 1 mm slice increased from 22% to 33%. The transparency of the 500 µm slice remained approximately twice as high as that of the 1 mm slice throughout the measurement, which may suggest a linear relationship between sample thickness and the increase in the transparency level after clearing with the CUBIC-R1 solution. In terms of the sample area, both tissue slices exhibited similar absolute expansion, with increases of 46% for the 500 µm slice and 39% for the 1 mm slice. This suggests that the absolute increase in the tissue area induced by CUBIC-R1 can be largely independent of tissue thickness. Figure 3(e) provides normalized plots of tissue transparency and area (min-max scaling) to facilitate comparison of the dynamics of these two parameters. When comparing the transparency plots, it can be observed that both exhibit a similar trend: a rapid initial increase in transparency followed by a gradual slowdown over time, resembling a logarithmic function. In contrast, the normalized area plots indicate that the expansion process is more dynamic for the thinner slice, with the largest sample area reached within the first 8 h of measurement. For the thicker slice, the maximum area is achieved after approximately 32 h. After reaching their maximum area, both slices gradually shrink over time, suggesting that the clearing duration should be carefully adjusted to the slice thickness in applications where the level of sample expansion is critical. When comparing the area and transparency plots, it can be observed that tissue expansion occurs earlier in the process, with most of the expansion completed within the first 5–10 h. During this period, only 20%–30% of the total transparency increase is achieved. This indicates that the proposed method may be useful for determining the optimal clearing time to achieve a desired balance between transparency and expansion, depending on the specific application and clearing solution employed.

The transparency plots presented in Fig. 3 show abrupt changes at the very start of the monitoring. The initiation of the tissue clearing process—triggered by the addition of a clearing agent to the slice—is highly dynamic, with early transparency fluctuations arising from the initial lipid release from the sample. These fluctuations manifest as caustic-like background perturbations, as illustrated in Fig. 3(a1), and the transparency values are shortly augmented, as calculations (see Fig. 2) are derived from the mean intensity values measured within the sample area (briefly higher due to caustics) and its immediate surroundings (briefly lower due to caustics).

To further validate the consistency of our data, we collected additional measurements for 4 samples with thicknesses of 500 µm (2 slices) and 1 mm (2 slices). The results, presented in Fig. 4, demonstrate that, while the final transparency values of the brain are influenced by the intrinsic properties of the tissue itself (which is heterogeneous in the case of the brain, particularly in terms of the white-to-gray matter ratio depending on the localization of the cutting plane and age of the animal), the overall trends in both the surface properties and sample transmittance over time are similar, reinforcing the reproducibility of our observations shown in Fig. 3.

Notably, the differences in transparency levels calculated by ClearAIM have a direct impact on the intensity of the signal obtained during subsequent fluorescence microscopic imaging, as shown in Fig. 5. We examined six cleared samples (after 1 day and 3 day incubation in CUBIC) to empirically validate whether the obtained outcomes translate into practical improvements when further working with the samples (fluorescence labeling and imaging). Interestingly, as shown in Fig. 5, 500-μm-thick sections derived from the same brain exhibited a depth-wise (using the same laser power) lower background signal caused by weaker tissue scattering after an extended clearing process (3 days) compared to a one-day clearing protocol. This finding aligns with our initially obtained ClearAIM numerical data regarding increased transparency.

Despite the rapid expansion of optical tissue clearing methodologies over the past decade—spanning well over 100 techniques and their optimizations—the field remains predominantly reliant on empirical observations by individual research groups^[1]. In this study, we introduced a new system for real-time monitoring of two critically important parameters for researchers employing optical clearing methods: (1) the level of transparency and (2) shape alterations, expressed as changes in the surface area of the cleared specimen.

Traditionally, the evaluation of transparency has relied heavily on the use of spectrophotometers or imaging with trinocular microscopes or professional cameras^[11,23]. These approaches, while widespread, have notable limitations. Spectrophotometric measurements necessitate the precise positioning of tissue within a cuvette, often requiring adhesive substances that not only influence measurement accuracy but are also incompatible with solvent-based clearing methods. Furthermore, comparing transmittance across techniques is complicated by their differing effects on tissue shape: some methods expand specimens by significant margins^[24], while others induce shrinkage^[25]. Such changes alter the surface area through which laser light passes, confounding measurements of light transmittance through the tissue versus the surrounding solution.

Macrophotography, on the other hand, faces challenges in maintaining consistent lighting and detection conditions, often necessitating professional photographic expertise—a resource inaccessible to many laboratories. Partial solutions to these limitations have been proposed by Woo et al.^[17] and Kim et al.^[18] However, their approaches fall short by precluding real-time assessment of transmittance and shape, instead relying on measurements at predefined, and thus subjective, intervals.

Our system addresses these limitations. By incorporating a precise segmentation algorithm, our method for evaluating specimen transparency remains unaffected by changes in specimen size. Additionally, the system continuously monitors specimen dimensions, providing researchers with critical yet frequently overlooked data essential for subsequent image analysis, such as estimating the expected size of segmented objects. Crucially, all data are generated in real time, even during extended, unattended experimental periods. The applicable high sampling rate enables precise monitoring of the clearing process even at very short intervals, which will be particularly valuable for assessing newly introduced ultra-fast clearing methods^[26,–28] that achieve tissue clearing in under 1 h. As demonstrated, this capability enables the correlation of transparency levels with the morphological effects of specific clearing methods. Researchers can, thus, make tailored decisions, such as determining whether a lower level of transparency suffices for obtaining valuable microscopic data while preserving a shape closer to the specimen’s original morphology. This flexibility supports more nuanced experimental designs and fosters a deeper understanding of the trade-offs inherent in various clearing techniques. Ultimately, the novelty of this approach lies not only in the implementation of precise segmentation algorithms or real-time monitoring but also in their synergistic integration, yielding a robust system capable of managing the complexities of diverse experimental contexts.

The method’s adaptability supports different sample thicknesses or tissue types, allowing for comparative studies that highlight subtle but important differences in clearing dynamics. Ultimately, this flexible and robust segmentation system paves the way for new research opportunities in imaging-based biological studies. It can easily be integrated into existing workflows that require tracking changes in tissue transparency, structure, or other properties. With its open-access availability, this method reduces barriers to adaptation and encourages broad community engagement.

An important technical innovation lies in how we maintain and update SAM segmentation prompts (positive and negative points) dynamically. Traditional segmentation pipelines often require manual adjustments or rely on fixed criteria that do not adapt well to non-uniform samples. In contrast, our method continuously refines the prompts, via the K-means algorithm, based on actual frame-to-frame changes observed in the data. For example, if the sample’s shape or texture shifts due to clearing-induced swelling or subtle rotations, the updated SAM prompts still guide the model accurately. This level of adaptability ensures that even if the optical properties of the sample change dramatically—transitioning from opaque to nearly transparent—the segmentation remains stable. As a result, the approach exhibits resilience against common challenges, such as brightness fluctuations or the presence of faint and irregular boundaries.

By providing a practical, powerful, and user-friendly tool, we expect our approach to become a valuable resource for scientists aiming to extract deeper insights from long-term imaging experiments, not only in the context of tissue clearing but also in related areas where samples can undergo spatial and optical transformations over time.

In aggregate, we introduced a new method for real-time monitoring of tissue clearing processes, leveraging a deep learning-based segmentation algorithm with a tailored frame-to-frame scheme. Demonstrated using mouse brain tissue sections, the method showcased its effectiveness through dynamic visualizations, including live transparency and size plots, achieving precise segmentation under CUBIC-induced alterations. Comparative analysis of tissue slices with varying thicknesses (500 μm and 1 mm) revealed distinct, sample-dependent clearing dynamics, emphasizing the quantitative power and reproducibility of the approach.

Our method provides three key advantages: (1) precise quantification of transparency levels and real-time tracking of morphological changes during clearing, (2) optimization of clearing timelines to balance tissue transparency and structural preservation, and (3) universal applicability to evaluate any tissue clearing protocol due to straightforward apparatus design.

By democratizing high-precision, real-time monitoring with open-access tools, this approach sets a new benchmark in tissue clearing, empowering laboratories to achieve reproducible, data-driven insights across diverse clearing methods and tissue types.