Tissue collection

Tissues submitted to pathology are typically classified by the collection procedure. Biopsies are generally incisional, meaning they involve removing only a small portion of tissue from a lesion to aid in diagnosis, and usually do not fully remove the abnormal area (radicality). Therefore, they are not curative, and final treatment decisions usually depend on the biopsy results. In contrast, excisional biopsies and surgical specimens (also called resections) involve removing the entire lesion or area of interest, providing enough tissue for definitive diagnosis and aiming for curative radicality with clear margins (the edge of healthy tissue around the removed lesion).

Side effects of both biopsy and surgical procedures can induce artefacts within the samples, such as cautery and crush artefacts (Fig 1A, and Table A and Figure A in Appendix S1).

The collected fresh samples are unsuitable for microscopic analysis, and once removed from the organism’s blood supply, they quickly start to degenerate. Hence, tissue samples must undergo a series of processes to stop degeneration and prepare them for microscopic examination.

Fixation

Fixation halts autolysis and degradation, stabilising both the tissue’s morphological architecture and its molecular components (e.g., proteins and nucleic acids) for downstream analyses. Furthermore, fixed tissue acquires a harder consistency and is thus easier to cut in further steps. Different fixatives can be used, but the universally recognised medium for tissue fixation is an aqueous solution of formaldehyde, which preserves the tissues and cells through the creation of cross-links among proteins [4]. Underfixation [5] and overfixation complicate tissue preparation, potentially affecting downstream analysis (e.g., immunohistochemistry (IHC) for predictive markers [6,7]). Additionally, underfixation corrupts tissue and cell morphology (Fig 1B).

Grossing

During the grossing step, the tissue samples are observed, described, and any alterations are noted and measured—macroscopic examination—and then selected (how many and which parts) to process for microscopic analysis—sampling—depending on the type of specimen and macroscopic findings. Inking can be used to mark certain parts of the tissue, commonly surgical margins, to make it recognisable on the final slide (Fig 1C). It is important to understand that for large specimens (e.g., an entire lung or colon), it is diagnostically not useful and economically infeasible to sample the entire organ. Therefore, sampling is strategic and guided by clinical and radiological findings to ensure the most relevant areas are captured.

Tissue processing and embedding

Tissue processing consists of tissue dehydration, through a series of immersions in alcohol and xylene, followed by paraffin wax infiltration. The tissue sample is then encased in a block of paraffin wax - tissue embedding - to stiffen and give it a homogeneous consistency, essential for microtome cutting.

It is important to note that during this stage, the spatial orientation of the tissue is fixed. Any auxiliary materials used during processing to secure the sample, such as biopsy sponges or papers, may be left around the tissue and be visible in the final WSI (Fig 1D).

Actionable insight: The orientation fixed during embedding determines the 2D plane of the final digital image; if a specific structure is not oriented correctly in the block, it may be visible suboptimally (or not at all) in the resulting WSI.

Microtome cutting

Once embedded in paraffin blocks, the tissue is sectioned using a microtome. Sections are typically cut at a thickness of 2–5 microns (µm), parallel to the block surface; sections that are too thick cause overlapping nuclei and loss of detail, while those too thin may lose structural integrity. These thin slices are then floated on a water bath and mounted onto glass slides.

Artefacts introduced during this stage can manifest as structural noise in the obtained digital slides after scanning (e.g., tissue folds, as in Fig 1E), complicating image analysis algorithms. Moreover, some of these can negatively impact the scanning phase by causing poor-quality (out-of-focus) images, ultimately leading to altered interpretation by AI algorithms [8].

Actionable insight: AI developers should treat these structural artifacts not merely as noise to be filtered, but as a source of data variability. Implementing heavy augmentation or developing automated quality control filters to exclude out-of-focus tiles is essential to ensure model reliability in clinical settings [8].

Staining

The unstained slices are mostly transparent, except for some intrinsic pigments. Consequently, staining is required to extract any significant morphological information and highlight different physical, chemical, or biomolecular characteristics of tissue. The most common routine staining technique using haematoxylin and eosin (H&E), which highlights nuclei (and other basophilic structures) in blue-purple, and cytoplasm and extracellular matrix proteins pink, respectively (Fig 2).

Download:

• PNG
  larger image
• TIFF
  original image

Fig 2. Examples of common stains: (A) H&E, (B) PAS, (C) Masson trichrome, (D) Alcian blue, (E) Giemsa, (F) Ki67 IHC.

Examples from the University of Leeds Virtual Pathology repository [9].

https://doi.org/10.1371/journal.pdig.0001148.g002

In addition to H&E (Fig 2A), several other stains are used, with the most common being the plethora of histochemical and immunohistochemical stains. Histochemical stains, such as periodic acid-Schiff (PAS), Masson trichrome, Alcian blue or Giemsa (Fig 2B–E, respectively), exploit some physical and chemical properties of tissues to stain them in a meaningful way, just like H&E. Immunohistochemical stains, on the other hand, leverage the very specific antibody-antigen interaction to highlight the presence of certain proteins. These interactions are visualised using chromogenic substrates (e.g., diaminobenzidine - DAB), producing a visible precipitate (brown, in the case of DAB), such as the example in Fig 2F.

Staining artefacts are mostly related to the variability of the chromatic properties of a stain (Fig 1F) across laboratories or even within the same laboratory. While modern laboratories often use fully automated staining platforms, these improve consistency but do not fully eliminate variability, since innumerable variables affect the final color, leading to both regional and temporal differences.

Actionable insight: Because staining variability is both regional and operator-dependent, developers must not rely solely on digital stain normalization. Models should be validated, and ideally trained, on multi-center data with diverse staining profiles rather than assuming a “standard” H&E appearance.

Coverslipping

Coverslipping serves the dual purpose of protecting the tissue on the slide from damage and of homogenising the light refractive properties of the tissue within the glass slide. Conventionally, the tissue placed on the glass slide is covered by a second, thinner piece of glass. A mounting medium, which polymerises and hardens, much like a glue, is put between the two pieces of glass and around the tissue. Some alternative methods, however, use a plastic film, a layer of polymer, or other materials as the coverslip. In high-throughput laboratories, this process is typically automated. However, artefacts can still emerge due to the presence of variability in mounting media, such as bubbles (Fig 1H) and regions of the tissue beyond the coverslip (thus not mounted), appearing as dark, opaque, and blurred regions in the digital image due to the refractive index mismatch.

Human vs. machine perception: It is important to note that while the laboratory steps described above can introduce artifacts, they rarely pose an insurmountable diagnostic limitation for a pathologist. In routine practice, if an artifact jeopardizes an assessment, in most (but not all) cases, the slide can be amended (e.g., re-coverslipped) or re-done (e.g., by recutting a new section from the paraffin block).

Actionable insight: For the AI developer, however, these artifacts represent significant computational limitations. Unlike a pathologist, an algorithm typically operates on the digital data provided without the ability to request a better sample. Therefore, what is a minor procedural inconvenience in a clinical workflow becomes a hard failure point for an AI model. Developers should frame artifact-heavy slides as a challenge of algorithmic robustness rather than a lack of diagnostic information in the underlying tissue.

Beyond histopathology

As the field matures, the principles outlined here for tissue sections will need to be adapted for cytopathology. AI developers should be aware that cytology samples - such as fine-needle aspirates or smears - present distinct challenges, including sparse cellularity, overlapping cells in different focal planes, and varying background artifacts. These topics represent a significant portion of the pathology workload and warrant a dedicated computational syllabus in the future.

Whole slide images

Whole slide imaging [15] is a cornerstone of digital pathology, enabling the visualisation of entire histological slides on a screen at high resolution. As the digital counterpart of a microscope, it is important to understand some of its technical specificities, particularly when designing digital pathology workflows or developing and deploying AI tools.

Pyramidal image structure.

WSI files are typically stored in a pyramidal image structure (Fig 3), in which multiple resolution levels are embedded within a single file. This design allows smooth zooming in and out (much like navigating digital maps) without needing to reload or decompress the entire image, since each level is downsampled from the base layer. Such a structure mimics the traditional experience of examining slides under a microscope, where a pathologist adjusts magnification dynamically to focus on areas of interest. Moreover, it also improves memory efficiency and processing speed, critical for both human experience and AI inference.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 3. Whole slide image (WSI) pyramidal structure.

Lower levels correspond to higher magnifications, thus higher resolutions, and vice versa. Example from The Cancer Genome Atlas (TCGA) dataset [16].

https://doi.org/10.1371/journal.pdig.0001148.g003

Advantages & challenges of slide digitisation

The usage of WSI represents a transformative step in modern diagnostic workflows, offering numerous advantages over traditional glass slides. It improves accessibility, collaboration, and long-term data preservation through remote access, AI-assisted analysis, and secure metadata storage. However, the digital transition introduces some challenges, such as image quality and standardisation, high storage and computing demands, and strict regulatory and compliance requirements for diagnostic use (Fig 4). Balancing both sides is essential, not only for the successful and scalable implementation of digital workflows in clinical practice but also for the development of efficient and useful AI tools.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 4. Advantages and challenges of digital transition in Pathology.

https://doi.org/10.1371/journal.pdig.0001148.g004

Basic diagnostic orientation

Before examining the slide, pathologists ground their interpretation in contextual information. The clinical context is fundamental to anchor the diagnostic hypothesis. Especially in nonneoplastic disease, i.e., inflammatory/autoimmune/infective or maladaptive disorders of organs and tissues, the same slide can be diagnostic for two different conditions, depending on clinical context (e.g., ulcerative colitis vs. segmental colitis associated with diverticula [SCAD]). Even in neoplastic disease, i.e., abnormal tissue growth resulting from uncontrolled cell growth, patient history (e.g., smoking or prior malignancy) drastically reshapes the expected spectrum of disease. Then the first step is to correlate what is seen in the slide with the macroscopic description and the material within the paraffin block, ensuring that the slide corresponds to the intended tissue (e.g., tumour, surgical margin, lymph node). Especially in fragmented specimens like biopsies, this confirmation step is essential to avoid misinterpretation. Tracking systems help document and verify specimen-slide correspondence, aiding both quality control and data integrity, which is also essential for AI development [19].

Actionable insight: Developers must recognize where AI algorithms can not currently be fitted into a workflow:

• Primary diagnosis without clinical context: AI should not be tasked with providing a definitive diagnosis for nonneoplastic diseases where the histology is identical, but the clinical history (e.g., drug use vs. infection) is the deciding factor.
• Incomplete sampling assessment: An AI model cannot confirm “clear margins” if the surgeon or pathologist did not sample the actual surgical edge during grossing.
• Fragmented specimens: In fragmented biopsies, AI models may struggle with orientation; developers should avoid projects requiring precise spatial localization unless some sort of orientation is provided to the model.

Understanding these “can not” scenarios prevents the development of “black-box” models that might make high-confidence predictions based on incomplete biological data.

Slide examination strategy: from low to high power

The slide examination proceeds from the overview, at the architectural level, in a low-power magnification, to the detail at the cellular level, in a high-power magnification (Fig 5). Using the 2× or 4× magnification, pathologists evaluate overall architecture. In tumours, they assess lesion boundaries (infiltrative/irregular vs pushing/well-defined), necrosis, cellularity gradients, and hotspot regions (e.g., for mitotic count). In nonneoplastic specimens, general patterns (fibrosis, inflammation, atrophy) are noted, and adequacy is assessed (e.g., number of glomeruli in a kidney biopsy or portal tracts in a liver core). Higher magnifications of 20× or 40× are used to interrogate specific structures, such as nuclear atypia, mitoses, inflammatory cell types, and inclusion bodies, that become evident (further described in the tumour glossary by Mandal S. et al. [3]). Immunohistochemistry or special stains may follow to assess specific biomarkers, but this level provides most morphologic discriminants.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 5. The pathologists’ examination strategy: different information is gathered at different magnifications.

Tissue-level aspects such as architecture are best assessed at low and very low power (A, B), whereas cell-level details require high power (C, D). Haematoxylin&Eosin (A–C) and diaminobenzidine immunohistochemistry for CD3 (D).

https://doi.org/10.1371/journal.pdig.0001148.g005

Common pitfalls and cognitive biases

The digital advent is transforming pathologists’ practice by enabling greater scalability, collaboration, and ultimately, computational analysis. However, despite such advancements, pathology assessment is susceptible to errors influenced by perception, experience, and cognitive bias. The digital medium does not eliminate these vulnerabilities, and in some cases, it may even exacerbate them. Thus, when developing reliable AI tools, it is crucial to recognise and understand the common pitfalls to maintain diagnostic accuracy and, ultimately, ensure patient safety. In a field where minor oversights can have significant consequences (e.g., missing a high-grade tumour or deciding on the wrong treatment), critical awareness is as important as technical innovation.

Segmentation tasks

Segmentation in digital pathology involves locating and labelling regions or objects within a WSI. This usually takes the form of semantic segmentation, usually used for tissue regions, or instance segmentation, usually used for specific types of cells (Fig 6). For these tasks, the ground truth can be precisely established: the pathologist can see and manually annotate a specific region or object on the slide, and the AI model aims to learn to do the same. However, such pixel segmentations are labour-intensive, thus many studies focus on tasks with public datasets, commonly associated to open challenges, such as CAMELYON [24] for breast cancer metastasis detection in lymph nodes, TIGER [25] for tumour and stroma segmentation (Fig 6B), GLAS [26] for colorectal gland segmentation, ICIAR BACH [27] for breast cancer grading or PANDA [28] for prostate cancer grading. Instance segmentation targets smaller, sparser elements, often requiring prolonged search to annotate, also leading to a focus on tasks with publicly available data. Examples include immune cells [25] (e.g., Fig 6C), mitotic figures for breast cancer prognosis [29], and nuclear segmentation (e.g., MoNuSeg [30]). Instance segmentation is also used for immunohistochemically stained slides, where AI can quantify protein expression (e.g., PD-L1 in lung cancer) more precisely than visual estimates [31].

Download:

• PNG
  larger image
• TIFF
  original image

Fig 6. (A) Breast tumour tissue, (B) with semantic segmentation of invasive tumour (light blue) and tumour-associated stroma (green), and (C) instance detection of immune cells (black boxes).

Example from the TIGER [25] challenge.

https://doi.org/10.1371/journal.pdig.0001148.g006

Common models used for the segmentation tasks include U-Net and its variants (nnU-Net [32], pathology-nnU-Net [33]), cell-specific architectures like HoVer-Net [34], HoverNext [35], and StarDist [36], and increasingly, transformer-based approaches (e.g., Swin Transformer [37]), sometimes enhanced by general or pathology-specific foundation models.

Due to the very large size of a typical WSI, slides must first be split into equally sized images, called patches or tiles, taken at the magnification considered best for the application and with a size optimised for computational resources and model design (Fig 7). The model is then applied to each individual tile, producing a patch segmentation that can be stitched back together to form a full slide segmentation, as a standard procedure implemented by several WSI-processing open-source libraries, such as the TIAToolbox [38], MONAI [39], or PathML [40].

Download:

• PNG
  larger image
• TIFF
  original image

Fig 7. (A) A cervical carcinoma WSI, (B) its tissue mask (light blue), and (C) a visualisation of where (D) tissue tiles (1024 × 1024 pixels) would be taken to feed a model.

Example from The Cancer Genome Atlas (TCGA) dataset [16].

https://doi.org/10.1371/journal.pdig.0001148.g007

Diagnosis and prognosis tasks

While segmentation focuses on localising specific morphologies, diagnosis and prognosis tasks predict slide- or patient-level information without explicitly delineating regions. An example of such a task would be the prediction of liver cell enlargement, i.e., hypertrophy, in drug safety studies, whose presence is more important to detect than its location for assessing drug toxicity. Diagnosis and prognosis tasks can also span multiple WSIs, such as when diagnosing the ISUP (International Society of Urological Pathology) grade of a prostate tumour to determine if systemic treatment is necessary. This requires the analysis of all the tissue from the prostate, which spans multiple slides.

Such an approach can also be used to predict patient prognosis, such as overall survival or treatment response. In these cases, since no pixel-level annotations are required, ground-truth labels can often be obtained from pathology or patient reports. Such weak labels make patch-wise segmentation-style approaches unreliable, given that only a small portion of pixels may be relevant. Instead, weakly supervised, e.g., multiple instance learning (MIL), is typically used: features are extracted from all patches using a large pre-trained network, aggregated across the WSI or patient, and then classified. Aggregation modules often employ attention or clustering. Popular pipelines include ABMIL [41], CLAM [42], TransMIL [43], and HIPT [44,45], while feature extraction increasingly relies on pathology foundation models such as CTransPath [46], Virchow2 [47], Phikon [48] and UNI [49], which offer strong generalisation for diverse diagnostic and prognostic tasks [50,51].

Building on this, the new frontier for computer-aided diagnosis and prognosis goes beyond simple WSI-only outputs [52]. It involves the development of visual question answering tools (e.g., SlideChat [53], PathChat [54]) and the automated generation of pathology reports and image captions (e.g., PRISM [55], HistoGPT [56]). The goal of this approach is to emulate the processing and synthesis of visual and contextual data involved in routine pathologic assessment and to create an end-to-end diagnostic system for pathologists to use.

Discovery tasks

While segmentation and diagnosis/prognosis tasks automate existing pathology workflows, discovery tasks aim to uncover novel links between pathology data and other biomedical modalities. Examples include cross-modal registration between WSIs and radiology images (e.g., prostate cancer [57]) and predicting genetic alterations directly from tissue morphology, first demonstrated in lung cancer [58] and later extended to multiple cancers [59], and preclinical studies [60]. These findings revealed previously unknown genotype–phenotype relationships and inspired the development of pathology-specific foundation models capable of robust performance across diverse tasks. Approaches vary: known-label problems such as mutation prediction often use weakly supervised methods with explainability tools (e.g., Grad-CAM [61]); cross-domain translation employs generative models (e.g., StyleGAN [62]); and foundation models typically adopt transformer architectures trained with self-supervised methods common in natural image analysis (e.g., DINOv2 [63]).

Supportive tasks

Developing AI models for core analytical tasks in digital pathology often requires substantial data preparation, much of which can be automated with supportive AI models. For example, WSIs contain large non-tissue regions that dilute the diagnostic signal and slow processing. These can be excluded using classical image processing (e.g., Otsu thresholding [64]), though such methods struggle with low tissue–background contrast, such as many IHC-stains and fatty tissue areas, prompting the development of AI-based tissue–background segmentation models [65,66].

Stain normalisation is another common preprocessing step. Variations in stain appearance, driven by site-specific protocols and preferences, can add noise without reflecting biological differences. Traditional approaches, such as the Macenko normalization [67], adjust stain concentrations but may fail with overlapping absorption spectra, as is common with H&E, leading to AI-based colour normalisation methods that are more robust across staining types [68,69].

Supportive AI models also assist with ground-truth generation for segmentation. Semi-automated approaches use pretrained models to produce initial segmentations, which are then corrected by pathologists. Pathologist-in-the-loop systems further refine models iteratively, reducing manual workload. Several open-source (e.g., VGG Image Annotator [70,71], SAM-for-Qupath [72]) and commercial tools (e.g., HALO [73], Discovery [74], and Aiforia Create [75]) employ this strategy.

Similarly, some models can also extract labels directly from patient or lab reports. Finally, supportive AI plays a role in deployment, such as WSI quality control. Routine-care slides may contain artefacts that compromise AI performance, and these specialised models can detect and exclude such slides (or even just affected regions within the slide) to maintain reliability in clinical use [76].

Challenges when developing AI for pathology

When developing AI models for pathology, several challenges are typically encountered. Some of these stem from the size and structure of WSIs, some are inherent to applying AI in a medical context. In Table 1, we summarise the most common ones for both sources. We then provide some information on the typical solutions and their limitations. Addressing these limitations is an active area of research. For example, to mitigate performance degradation caused by tiling WSIs for slide- or patient-level tasks, several MIL approaches have proposed sophisticated tile aggregation methods that help restore some of the context lost. To tackle the inherent data variability of the slides, there is an active effort to choose models that are invariant to such batch effects, rather than relying on additional preprocessing and quality control. One common strategy involves mapping the feature space of a dataset to a lower dimension (e.g., using t-SNE) and examining whether there are clusters that correspond to unintended sources of variability (e.g., centre, preparation method, scanner). Together, these efforts aim to enhance the reliability and usability of AI models developed for pathology, moving these novel technologies a step closer towards translation.

Download:

• PNG
  larger image
• TIFF
  original image

Table 1. Typical challenges when developing AI for pathology.

https://doi.org/10.1371/journal.pdig.0001148.t001