Spatial Phenotyping in Digital Pathology: Defining a multiplex IF image analysis workflow to expedite translational and clinical R&D
Recent advances in multiplex staining and digital pathology have paved the way for enhanced spatial profiling. Within quantitative digital pathology, the exploration of cell spatial phenotyping has emerged as a powerful tool for delving deeper into the intricate landscapes of tissue sections.
In this article, we dive into the world of cell spatial phenotyping, outlining its significance and the essential steps in generating spatial data from multiplex immunofluorescence (mIF) stained tissue sections. We will focus on the application of a standardized image analysis workflow that can be robustly applied to studies incorporating different marker panels and tissues, and how new innovations are integrated to continue to support efficiencies moving forward.
Image Analysis Workflow
OracleBio’s mIF image analysis workflow covers 4 key areas, including:
Each step is an important precursor to the next step regards the quality and efficiency of the final data.
Let’s look through these steps in more detail below.
Image Quality Control (QC)
Each image undergoes detailed quality checks. Factors including 1) image scanning quality, 2) tissue integrity, and 3) stain quality are assessed.
For multiplex images, each channel of the image is visually assessed by our image analysis scientists for these factors following our in-house standard operating procedure (SOP).
An overall pass/fail assessment is recorded and returned to the client, with only those images that pass continuing to the analysis phase. Images that fail can either be rescanned or restained at their discretion and re-entered into the study.
Algorithm Development
Algorithm development forms an integral part of the workflow, with the key aim to develop Apps that provide both precision; to capture the accuracy and detail of required study read-outs, and robustness; to work across different tissue and staining heterogeneities present in study samples or between study batches.??
The increasing access to AI Deep Learning networks available across a range of commercial (Visiopharm, Halo) and open source (Q-path) software has aided in the performance of Apps to deal with the above factors, but Deep Learning approaches need to be used correctly in the context of each study remit. For example, this can include choosing the correct neural network, learning rate, freeze depth, and resolution in line with the features to be detected in the training regions.
For a given study, the workflow revolves around two primary components: tissue classification and cell analysis.
Figure 2 below shows examples of an AI App (DeepLab3, Visiopharm) developed to classify tumor, stroma, glass, and artefact regions of interest (ROI) across mIF-stained TMA cores (courtesy of Akoya Biosciences). Although the cores contained a PanCK marker (orange), the Deep Learning network helped improve tissue classification across different cores containing variant PanCK expression, stain intensity, and tumor morphology heterogeneity.
Accurate cell segmentation is pivotal for meaningful phenotypic interpretation. Cells of varying lineages, shapes, and sizes populate the tumor microenvironment (TME) of cancer tissue.
Our approach involves training Deep Learning algorithms to recognize distinct cell lineages associated with the underlying pathology. For example, the TME can include macrophages, immune cells, tumor cells, and stroma cells. By sequentially annotating and training the algorithm for specific lineages, we can achieve precise detection and differentiation. This lineage-focused segmentation facilitates improved phenotyping and co-localisation staining of the underlying cell types present in the TME, enabling robust interpretation of cellular interactions.
领英推荐
It is also important to achieve optimal thresholding for stain positivity. Choosing a single global threshold can be challenging in relation to any observed stain variance across samples when analyzing large study batches. For cell types used in the Deep Learning training, ensuring cell annotation training regions are captured from tissue with different signal-to-noise ratio, or with any non-specific or autofluorescence staining, will help to generalize the network when applied to samples with variant staining present.
Global thresholds are manually chosen for remaining markers and phenotypes created as required for the study read-outs. Thresholds and phenotypes are checked over various samples to ensure accurate cell phenotyping in line with the provided study samples.
Algorithm validation for specific studies is also an integral part of this phase. Here both classifier and cell analysis algorithms can be validated against a pathologist ground truth read. OracleBio’s SOPs cover a DICE coefficient scoring method for classifier validation and a correlation analysis method for cell analysis validation. Note, that for multiplex images, each marker or channel will require its own validation.
Image Processing
Once developed, and validated, algorithms are put to work processing images and generating data.
Here, our cutting-edge, purpose-built cloud IT infrastructure is essential in delivering efficient processing of large mIF images that have numerous data read-outs.
At OracleBio, both our HALO and Visiopharm software is hosted within this cloud environment (AWS), leveraging unlimited storage and access to high-performance, scalable GPUs and CPUs on demand.
This enables us to batch analyse large numbers of images in parallel to generate data extremely efficiently.
We further use web-based portals including HALO Link and our ‘OBserver’ platform (proprietary Visiopharm web viewer) to enable clients to log in remotely to review analysis we have performed. This also acts as a collaborative tool for tissue annotation generation and algorithm development review during the study.
Data generated per image is exported, rendered, and QC’d before being returned to the client as a data package with the requested read-outs per ROI highlighted. A copy of all ‘raw data’ files directly from the software is also provided.
Data Management & Spatial Insights
Spatial analysis of digital pathology data can provide further biological or therapeutic response insights on top of changes in cell phenotype numbers/counts. Examples of spatial data include infiltration analysis, cell-to-cell proximity, distance relationships, cluster analysis, and neighbourhood analysis.
Cell object data files, exportable from image analysis software, contain a wealth of information – mean stain intensity, XY coordinates, phenotype labels, and more. Through scripting languages like MATLAB and Python, we can explore advanced spatial profiling techniques.
By interrogating data from across specific ROI, we can unveil insights into the spatial tendencies of cells. Such analyses expose patterns of dispersion, clustering, or randomness, vital for understanding cellular interactions within a tissue microenvironment. Local and global tissue heterogeneity evaluations provide a lens into the organization of cellular neighbourhoods. By gauging the ability of cellular regions to form neighbourhoods, we gain insights that might correlate with prognosis or therapeutic response. This knowledge can foster the development of companion diagnostics, enabling precise patient stratification.
See further data in our poster presented at AACR 2023.
Summary
Quantitative digital pathology within R&D remains a highly dynamic area with constant innovations in imaging, IT infrastructure, AI, analysis software, and data management processes all impacting the quality and efficiency of final data generation. Furthermore, spatial phenotyping using mIF techniques is shedding light on the complex cellular landscapes of tissue sections, offering insights beyond traditional cell counts.
Moving forward, the fusion of deep learning into AI image analysis algorithms and the rise of scalable cloud computing will further bolster the efficiency of spatial data analysis, and collectively, these technologies synergize to enable the extraction of comprehensive cell and spatial data from complex tissue pathologies.
Critical to this is the continuing interactions between anatomical pathologists working alongside image analysts as part of a standardized workflow will provide the best platform for incorporating new innovative approaches, thereby delivering on the potential of spatial biology to help identify improved therapeutic approaches and drive personalized medicine.
If you’d like to learn more about spatial phenotyping and how OracleBio can support your R&D, get in touch with our experts.