- Tissue image processing and registration are important computational steps in spatial transcriptomics and spatial omics. These procedures prepare tissue images for molecular analysis and ensure that gene-expression measurements, detected transcripts, and cellular features are accurately associated with their original positions in the tissue. Spatial omics technologies generate molecular information together with microscopy images, histological sections, fluorescence signals, or other imaging data. However, raw images may contain noise, uneven illumination, tissue damage, distortions, background signals, and differences between imaging channels or experimental cycles. Image processing improves image quality, while image registration aligns images so that corresponding tissue structures and molecular measurements can be compared accurately.
- Tissue imaging commonly begins with the preparation of thin tissue sections. Samples may be fixed, embedded, sectioned, stained, and mounted on slides or specialized spatial capture platforms. Hematoxylin and eosin (H&E) staining is widely used to visualize tissue morphology, including nuclei, cytoplasm, extracellular structures, and tissue boundaries. Immunohistochemistry and immunofluorescence provide information about selected proteins and cellular markers, while fluorescence microscopy can detect labelled RNA probes and other molecular signals. Spatial transcriptomics platforms such as 10x Genomics Visium combine tissue images with spatially barcoded capture areas, whereas imaging-based platforms such as Xenium, CosMx, MERFISH, and seqFISH generate high-resolution molecular images that require computational processing.
- The first stage of tissue image processing involves image acquisition and quality assessment. Microscopy systems may generate bright-field, fluorescence, confocal, or other forms of digital images. Important acquisition parameters include image resolution, pixel size, magnification, exposure time, illumination intensity, focus, optical sectioning, and the number of imaging channels. Images should be acquired with consistent settings whenever samples are intended for direct comparison. Poor focus, uneven illumination, overexposure, underexposure, low signal intensity, and tissue damage can reduce the accuracy of subsequent computational analysis. Metadata describing the imaging system and acquisition settings should be retained to support reproducibility.
- Image preprocessing is performed to improve the quality of raw images before registration and molecular analysis. One common procedure is background correction, which removes unwanted signals caused by autofluorescence, nonspecific staining, optical noise, or the imaging environment. Background subtraction can improve the contrast between true molecular signals and surrounding tissue. Illumination correction addresses uneven light distribution across an image, which may cause some regions to appear brighter or darker than others. Flat-field correction is commonly used to compensate for spatial variations in illumination and imaging-system response.
- Image denoising is another important preprocessing step. Microscopy images may contain random intensity variations, detector noise, and other artefacts that interfere with the identification of cells and molecular signals. Filtering methods such as Gaussian filtering, median filtering, and non-local means can reduce noise while preserving important image structures. More advanced approaches use machine learning and deep learning to remove noise and enhance image quality. However, excessive filtering can remove small cellular structures or weak fluorescent signals, so preprocessing parameters must be selected carefully.
- Image enhancement improves the visibility of tissue structures and molecular signals. Contrast adjustment, histogram equalization, and adaptive histogram equalization can help distinguish nuclei, cytoplasm, tissue boundaries, and fluorescent signals. In fluorescence imaging, signal intensity may vary considerably between channels or experimental cycles. Intensity normalization can help make images more comparable, although biological differences in signal abundance should not be unintentionally removed. Image enhancement should therefore improve visualization without introducing artificial structures or changing the underlying quantitative measurements.
- Image segmentation is used to identify meaningful structures within tissue images. It may involve separating tissue from background, identifying tissue regions, detecting nuclei, defining cell boundaries, and distinguishing different anatomical compartments. Threshold-based segmentation, watershed algorithms, edge detection, and machine-learning methods are commonly used. Deep-learning tools such as Cellpose and StarDist can identify nuclei and cells in complex microscopy images. CellProfiler supports automated image-analysis workflows, while QuPath provides tools for tissue annotation, cellular detection, and quantitative histological analysis. Accurate segmentation is especially important in single-cell-resolution spatial transcriptomics because detected RNA molecules must be assigned to the correct cells.
- Image registration is the process of aligning two or more images so that corresponding structures occupy the same spatial positions. In spatial omics, registration may be required to align an H&E-stained tissue image with a spatial transcriptomics capture image, combine fluorescence channels, align images from consecutive tissue sections, or match images obtained during different rounds of multiplexed imaging. Registration allows molecular measurements to be interpreted in relation to tissue morphology and other biological features. Without accurate alignment, spatial coordinates may be incorrectly assigned, leading to errors in tissue-region identification, cell localization, and downstream spatial analysis.
- Image registration generally involves identifying corresponding features or landmarks between images. These may include tissue boundaries, nuclei, blood vessels, glands, cell clusters, and other recognizable structures. The image used as the spatial reference is called the fixed image, while the image being aligned is called the moving image. A transformation is calculated to map coordinates from the moving image to the fixed image. Depending on the type of image variation, the transformation may be rigid, affine, or non-rigid.
- Rigid registration allows translation and rotation without changing the shape or size of the image. It is useful when images differ mainly in their position or orientation. Affine registration additionally permits scaling and shearing, making it suitable for images that have undergone differences in magnification or geometric distortion. Non-rigid registration allows local deformations and is useful when tissue sections have been stretched, compressed, folded, or distorted during preparation. Non-rigid transformations are more flexible but can also introduce artificial distortions if not appropriately constrained.
- A mathematical representation of image registration can be expressed as:
- x’ = T(x)
- Here, (\mathbf{x}) represents the original coordinate of a pixel or molecular observation, (T) represents the transformation function, and (\mathbf{x}’) represents its new coordinate in the registered image. The transformation may involve translation, rotation, scaling, shearing, or local deformation, depending on the registration method.
- Registration algorithms use different strategies to determine the best transformation. Feature-based registration identifies recognizable structures or keypoints in both images and estimates their correspondence. Intensity-based registration compares image intensity patterns and searches for a transformation that maximizes similarity. Landmark-based registration uses manually or automatically selected corresponding points. Deformable registration uses local transformations to align structures that cannot be accurately matched through a single global transformation.
- Image similarity can be evaluated using measures such as mean squared error, normalized cross-correlation, mutual information, and structural similarity. Normalized cross-correlation is useful when corresponding images have similar intensity patterns, while mutual information can be valuable when images come from different imaging modalities with different intensity distributions. Landmark-based evaluation can also assess how accurately corresponding tissue structures are aligned. The choice of evaluation measure depends on image characteristics and the biological structures being compared.
- In multiplexed imaging technologies such as MERFISH and seqFISH, image registration is particularly important because multiple rounds of fluorescence imaging are performed to identify different RNA molecules. The tissue may shift slightly between imaging cycles because of stage movement, changes in focus, tissue deformation, or other experimental factors. Registration aligns images from different cycles so that fluorescent signals can be associated with the same physical locations. Accurate registration is essential for decoding molecular barcodes and identifying individual transcripts. Poor alignment may cause false transcript identification, missed signals, or incorrect spatial localization.
- Registration is also important when integrating spatial transcriptomics with histological images. In sequencing-based platforms such as Visium, the tissue image must be aligned with the spatial capture grid so that gene-expression measurements can be associated with the correct tissue regions. Researchers may annotate anatomical structures such as tumour areas, stromal regions, blood vessels, glands, and immune-cell-rich compartments. These annotations can then be combined with spatial gene-expression profiles to investigate molecular differences between tissue regions. In imaging-based platforms, registration may connect molecular signals with H&E images, immunofluorescence channels, or other microscopy measurements.
- Several computational tools support tissue image processing and registration. Fiji/ImageJ provides image filtering, enhancement, measurement, and registration plugins. CellProfiler supports automated image processing and quantitative cellular analysis. QuPath is widely used for histological image analysis, tissue annotation, and cell detection. OpenCV provides computer-vision algorithms for image transformation, feature detection, and registration. Python libraries such as scikit-image, NumPy, and SciPy support image processing and numerical analysis. SimpleITK and elastix provide registration capabilities, including methods for medical and biological images. Advanced workflows may also use deep-learning frameworks to improve segmentation, registration, and image reconstruction.
- Spatial omics software packages provide additional support for integrating images with molecular data. Space Ranger processes 10x Genomics Visium data and connects spatial expression measurements with tissue images. Xenium Ranger supports processing of Xenium datasets. Seurat, Scanpy, Squidpy, Giotto, and SpatialExperiment support downstream spatial analysis and visualization. These tools can be used to explore gene-expression patterns in relation to tissue structures, cellular distributions, and spatial coordinates. Image-processing and registration tools are often used alongside these packages rather than replacing their platform-specific processing pipelines.
- After image processing and registration, quality control is necessary to determine whether the resulting images and spatial coordinates are reliable. Important quality indicators include image sharpness, signal-to-background ratio, tissue coverage, registration accuracy, alignment of anatomical landmarks, cell segmentation quality, and consistency between imaging channels. Researchers should visually inspect registered images and examine whether important tissue structures are correctly aligned. Registration errors can be especially problematic near tissue boundaries, damaged regions, or areas with substantial deformation. Poor-quality regions may need to be excluded or analyzed separately.
- Tissue image processing and registration have important applications in cancer research, neuroscience, developmental biology, immunology, infectious disease research, and animal genomics. In cancer research, image registration connects molecular profiles with tumour morphology, allowing researchers to investigate tumour regions, immune infiltration, and stromal organization. In neuroscience, it supports the mapping of gene expression to specific brain structures and anatomical regions. In developmental biology, images from different developmental stages can be aligned to investigate tissue growth and molecular changes. In animal genomics and breeding research, these methods can help relate gene-expression profiles to histological features in muscle, mammary glands, reproductive organs, liver, and immune tissues. This can support investigations of muscle development, meat quality, milk production, fertility, disease resistance, and tissue responses to environmental conditions.
- Despite their importance, tissue image processing and registration have several limitations. Tissue preparation may introduce deformation, tearing, folding, and loss of cellular material. Fluorescence imaging can be affected by autofluorescence, photobleaching, and signal variation. Images obtained from different staining methods or imaging platforms may have different resolutions and intensity characteristics, making registration difficult. Non-rigid registration may improve alignment but can also distort tissue geometry if the transformation is too flexible. Automated segmentation and registration algorithms may perform differently across tissues, species, and experimental conditions. Therefore, computational results should be assessed through visual inspection, quantitative metrics, and biological validation.
- Overall, tissue image processing and registration provide the visual and spatial foundation for reliable spatial omics analysis. Image preprocessing improves image quality, segmentation identifies cells and tissue structures, and registration aligns molecular measurements with the correct anatomical locations. Combining these procedures with sequencing-based and imaging-based transcriptomics enables researchers to connect gene expression with tissue organization and cellular relationships. Accurate image processing and registration are therefore essential for extracting meaningful biological information from complex spatial datasets.