Utilizing AI to improve analysis accuracy and turnaround time for Preclinical tumor model IHC studies
29 September 2022
By Mark Laurie

Article Synopsis: In this article, we highlight a previous webinar case study where we utilized AI deep learning techniques to improve the accuracy and turn-around time for quantification of cytotoxic T-cell (CD8 marker), macrophage (F4/80 marker), and vasculature (CD31 marker) IHC staining in viable tumor across n=10 different syngeneic tumor models.

In pre-clinical syngeneic tumor model immunohistochemistry (IHC) studies, a standard quantitative image analysis workflow involves segmentation of the viable tumor region of interest (ROI) from other components within the tissue, such as necrosis, artefacts and non-tumor host tissue, before quantifying the IHC stained cells or biomarkers of interest within the viable tumor ROI.

Depending on the number of samples in the study, the range or tumor models within the sample set, or the variable levels of specific or non-specific staining present across samples, the process of accurately segmenting ROI and detecting cells/biomarker staining can be challenging and time-consuming.

The study contained n=176 samples, stained across 3 serial sections (1 per IHC marker), providing a total of n=528 whole slide images. All samples were stained via single chromogenic IHC using Di-amino benzidine (DAB) chromogen (brown) to highlight positive staining. Nuclei were counterstained blue with haematoxylin. All image analysis was performed using Indica Labs HALO and HALO AI software.

The workflow was split into 3 steps:

Step 1: QC and Tissue Detection

An automated ‘tissue detect’ app was developed in HALO AI using the MiniNet neural network with 3 classes: whole tissue, gross artefacts, and glass.

This created an automatic annotation layer on each section for whole tissue, excluding gross artefacts across different models and IHC stains.

Figure 1: Examples of automated tissue detection using the Deep Learning ‘tissue detect’ app. Whole tissue outline (purple solid line); artefacts (purple dashed line).

Step 2: AI segmentation of viable tumor ROI

A Deep Learning ‘classifier’ app was then developed to separate viable tumor from host tissue and Necrosis ROI across all 3 marker image sets.

Representative areas of tumor, host tissue, Necrosis and artefacts were digitally annotated across up to 60 images covering 10 different models and IHC stains.

Annotations were trained using the HALO AI DenseNet v2 neural network at 1.5µm/pixel resolution and to ~60,000 iterations.

Figure 2: Examples of viable tumor AI classifier mark-up training regions. Viable tumor (green outline), host tissue (yellow outline), Necrosis (red outline), artefacts (blue outline).

Figure 3: Examples of viable tumor AI classifier output using the HALO AI DenseNet v2 neural network. Viable tumor (green outline), host tissue (yellow outline), Necrosis (red outline), artefacts (blue outline).

Step 3: Immune cell and vasculature IHC stain detection

For accurate CD8 immune cell detection, a nuclei AI segmentation app was developed in HALO to form cell objects based on the Haematoxylin staining within viable tumor ROI.

Formed cell objects were then classified as positive or negative based on a defined CD8 IHC DAB optical density threshold.

Figure 4A [Above]: CD8 cell quantification in Viable Tumor ROI

For F4/80 quantification, due to the distribution of the IHC staining not being directly associated with a nucleus, an area quantification app was developed to detect the area of F4/80 IHC DAB positive staining within the viable tumor ROI. 

Figure 4B [Above]. F4/80 area quantification in viable tumor ROI

For accurate CD31 vasculature quantification, a Deep Learning classifier was initially developed to segment vasculature from non-vasculature regions within the viable tumor ROI. This allowed for areas of non-specific, non-vasculature, DAB staining within viable tumor to be avoided.

Once vessels were detected accurately by the Deep Learning classifier, a CD31 area quantification app was then run within the ‘vessel ROI’, using a defined CD31 IHC DAB optical density threshold.

Figure 5. CD31 IHC stain area analysis in viable tumor ROI

In Summary

We applied several AI Deep Learning approaches to support the accurate and efficient execution of the study workflow, enabling us to expedite the delivery of detailed data back to our client.

The AI Deep Learning steps supported:

  • Reduced sample quality control time by automatically detecting and excluding tissue artefacts
  • Reduced sample annotation time by automatically segmenting host tissue from tumor tissue
  • Improved accuracy in segmentation of viable tumor across 10 different syngeneic models
  • Improved accuracy in detection of CD8 and CD31 specific IHC marker staining

OracleBio’s image analysis workflow for quantification of pre-clinical tumor model tissues standardly incorporates AI-powered apps run in a GPU-scalable cloud computing environment to ensure efficient generation and return of data.

Supported by our internal clinical pathologists, our image analysis scientists can develop apps tailored to specific studies with the option for our clients to own the apps at the end of the study, allowing them to be integrated back into their internal digital pathology workflows.

We have extensive experience in AI app development and quantification of IHC, ISH, mIF, and high-plex stained samples.

If you’d like to find out more about this approach or any of the services we provide, please get in touch with our experts at oraclebio.com/contact.

James Going

About the author: 

Lorcan Sherry, PhD. – Chief Scientific Officer at OracleBio

As co-founder and Chief Scientific Officer at OracleBio, Lorcan has helped establish the company as a market leader within its field, supporting the digital pathology programmes of major Pharma and Biotech companies around the globe.

Find out more about OracleBio's Pathology Expertise

OracleBio's Quantitative Digital Pathology Services

Stay up to date with OracleBio

Sign up to our mailing list to stay informed about news, webinars, events and more