Fig. 1. Light-sheet fluorescence microscopy^[1]. Left: archetypal light-sheet microscope, and paired orthogonal lightpaths provide plane-wise illumination (blue) and wide-field fluorescence detection (green); right: optical sectioning by selective illumination of a single plane

Fig. 2. Overview of SPIM^[4]. (a) Schematic of the sample chamber; (b) a medaka embryo imaged with SPIM; (c) time-lapse imaging of Drosophila melanogaster embryogenesis

Fig. 3. Parallelization of light-sheet generation^[1]. SPIM illuminates and captures fluorescence from the entire FOV simultaneously, whereas mSPIM reduces striped artifacts by pivoting the light-sheet about its center. DSLM produces a virtual light-sheet by time-sharing the beam, with fluorescence arising only from the illuminated strip at any given time. To maintain identical SNRs, DSLM requires higher peak intensities (I_peak) as the FOV size increases (along the scanning axis) relative to the light-sheet thickness

Fig. 4. Whole-brain single-cell-resolution imaging^[18-19]. (a) Tissue-clearing methods allow whole-brain profiling of cells; (b) images of an adult mouse brain obtained by hydrophilic tissue-clearing and custom-made high-resolution light-sheet microscopy; (c) 3D single-cell-resolution mouse brain atlas

Fig. 5. Application of light-sheet microscopy in pathological tissue samples. (a) False-color simulated H&E stained volumetric image of a human prostate core-needle biopsy sample and its orthogonal 2D cross-section, with magnified views of the region of interest at different depths on the right^[12]; (b) visualization of drug delivery throughout the entire tumor mass of the BT-474 human breast cancer cell line^[27]; (c) large-scale fluorescence imaging of thick human brain tissue section, with magnified view showing amyloid-rich areas^[31]; (d) label-free imaging results of the back of human tongue in vivo, H&E stained images of the filiform and fungiform papillae tissues in the tongue, desktop and micro MediSCAPE imaging results of the area of ​​interest^[22]

Fig. 6. Virtual staining results. (a) Pathological tissues (skin, kidney, prostate cancer, basal layer of the epidermis, kidney tubules, and prostate glands) stained with FalseColor-Python achieve different styles of results by adjusting staining parameters^[37]; (b) comparison of lymph node before and after virtual staining, the left side shows visualization stained with H&E fluorescence analog, while the right side shows visualization with FalseColor-Python virtual H&E staining^[23]; (c) using Cycle-GAN to transform pathological images from H&E staining style to IHC staining style accurately predicts the expression distribution of Ki-67^[40]

Fig. 7. Framework of pathological image analysis based on deep learning, including annotation, training, and inference methodologies^[41]. (a) Pathologists use computer-assisted annotation software to label PD-L1 status based on patients' H&E-stained and corresponding IHC-stained tissue microarray images; (b) using five-fold cross-validation, the CNN model is trained on H&E images from the training set, and prediction scores for each patient are generated on the validation and test sets, with the maximum score compared to the pathologist's annotation results for statistical analysis

Fig. 8. TriPath computational workflow^[47]. (a) 3D imaging modalities can capture high-resolution volumetric images of tissue specimens; (b) TriPath first separates the volumetric image of tissue from the background, and then divides the original volume image into small patches; (c) patches are processed with a pretrained feature encoder network of choice (3D CNN or 3D Vision Transformer)

Fig. 9. Introduction to 3D pathological analysis. (a) 2D cross-sections of false-colored H&E analogue images, synthetic-CK8 IHC images, and gland-segmentation masks based on the synthetic-CK8 images (left three columns), and 3D renderings of gland segmentations for benign and cancerous regions (right) ^[48]; (b) a micrometastasis (<2 mm) in the simulated histology images (top), deep 3D image of macrometastasis (>2 mm) (bottom left), and high-resolution image of the interface between the metastasis and benign tissue (bottom right) ^[23]; (c) single cells were automatically quantified using the Amira and MATLAB software packages, the level of vimentin in each individual cell of the tumour was assessed, and the vimentin expression was then rendered in 3D^[53]

Fig. 10. Overview of large pathological analysis model. (a) Overview of UNI^[56]; (b) overview of Prov-GigaPath^[57]; (c) overview of Virchow^[58]

Fig. 11. Core components of MLLM

Fig. 12. Curation of instruction-following dataset and PathChat overview^[66]. (a) The currently largest instruction fine-tuning dataset specifically for the domain of pathology, consisting of 456916 instructions and corresponding responses covering various formats, and this assistant is capable of reasoning with both visual and natural language inputs; (b) training process of PathChat

Fig. 13. An illustration of overall framework of PathAsst ^[67] . The multimodal MLLM training encompasses the training processes of both PathCLIP and PathAsst, as well as the construction of a paper embedding database. The tool-augmented MLLM inference details the process of PathAsst utilizing various tools to enhance the quality of its generated outputs

Fig. 14. Overview of M3D-LaMed model^[68]. (a) 3D image encoder is pre-trained by cross-modal contrastive learning loss with image‒text pairs, performing image‒text retrieval; (b) in the M3D-LaMed model, 3D medical images are fed into a pre-trained 3D image encoder and an effective 3D spatial pooling perceiver to produce refined embeddings inserted into LLM