Improving diagnostic efficiencies in histology with artificial intelligence and digital scoring

Digital analysis and artificial intelligence (AI) are not novel in anatomic pathology and histology, but the technology continues to evolve as it supports clinical trial enrollment as well as routine work in pathology.

Paul Mesange, PhD, Principal Scientist, Histology, recently shared how Labcorp Drug Development (formerly known as Covance) has implemented and validated AI-based digital software to help analyze and identify sub-types of breast cancer. Learn how these technologies are designed to improve diagnostic efficiencies, inform prognoses and guide treatment decisions for patients.

Understanding the standard process for informing a diagnosis  

The diagnostic process for breast cancer typically starts with a clinical examination followed by imaging and a pathologic assessment. A tissue biopsy is collected and is then embedded into a paraffin block and sectioned onto slides. From here, histotechnologists can perform different histochemical staining, such as fluorescent in situ hybridization (FISH), in situ hybridization (ISH) and immunohistochemistry (IHC) and run biomarker assays. These assays can help identify breast cancer subtypes by measuring, for example, the expression of markers, such as HER2, ER, PgR and Ki67. To ensure consistent results, these robust processes use the same platforms and procedures across the Labcorp network.

At the stage of analysis and interpretation, a board-certified pathologist views each slide to provide an evaluation. Results are then sent to a database, and, if applicable, the pre- and post-treatment tissue can be compared to help assess and predict therapeutic efficacy in specific disease states or in specific patient populations.

Incorporating digital analysis and artificial intelligence

Typically, pathologists can spend 5-15 minutes reviewing each slide, depending on the scoring needed. With many duplex assays that include two markers, the pathologist’s review time can range from 10-30 minutes per slide. Other multiplex assays with fluorescence, for example, can have six or more markers, which adds complexity to the review process.  

To assist the anatomic pathology and histology process—and improve efficiencies in the review and diagnostic interpretation of assays—digital analysis and AI can precede the pathologists’ reviews. Instead of going directly to the pathologist, slides are first analyzed through a digital software. AI technologies can assist analyses by defining phenotypes, which can then be reviewed by the pathologist before results go into the database.

Consider the data cited in 2011, where an estimated “1 million prostate biopsies were performed annually among Medicare beneficiaries.” Of those biopsies, approximately 75-80% are found to be non-cancerous, implying that pathologists studying these samples are mostly reviewing benign tissue, as described by Madabhushi et al.

If AI and digital analysis can easily and accurately distinguish benign tissue, pathologists can devote more time to studying unknown or potentially cancerous tissues.

How AI-based, assisted-image analysis supports identification

At a visual level, the following image shows how AI-based assisted-image analysis supports pathologists. In this image of breast cancer tissue, a pathologist is attempting to identify Ki67 expression. The left side shows the IHC staining of a patient’s sample, where the brown nuclei are considered positive and blue ones are considered negative. The first step of analysis is to identify the region of interest (ROI) to determine the area of tumor cells. The blue dotted line was identified by the AI algorithm.

Zooming in on the tumor region as shown below, the algorithm first identifies the tumor nuclei that are present (image on the left). The middle image displays outlines of the nuclei identified by the algorithm, while the right image shows the nuclei outlined in red, which are classified as Ki67-postive. This information can help the pathologist better understand the tumor area, the percentage of Ki67-positive cells and its distribution.

Validating results to ensure concordance between pathologists and algorithms

Labcorp has run in-depth validation, following CAP/CLIA recommendations, to compare its algorithms’ results with pathologists’ results to determine if AI can match or even outperform the human eye. With a set of individual samples, the percentage of Ki67-positive cells was separately tallied by the pathologist and the algorithm, and these counts were then averaged. If both the pathologist’s count and the algorithm’s count were less than 10 percentage points from the mean, it was considered acceptable. This validation has been performed for HER2, PgR and ER as well as numerous samples as Labcorp uses the technology to support sponsors’ clinical trials.

Looking ahead to increase diagnostic efficiencies in other cancer types

Labcorp currently has four algorithms in production for breast cancer: Ki67, ER, PgR and BCL-2. In addition, 10 algorithms have been developed for multiplex IHC, which are traditionally challenging to read under the microscope.

To further support routine pathologist work, the team is developing a tumor-detection application to support analysis of prostate biopsies and bladder cancers. Labcorp is also extending their technologies to other clinical indications and assisting pharma sponsors with their internally developed AI applications, which can also be transferred to Labcorp to help advance their clinical trials and ultimately help reach faster diagnoses for patients.

Read more about our immunohistochemistry services.