Deep Learning blog

Optimising Quality Control in clinical trials by automating quality tissue assessment

By David Wallis, Thomas Le Meur - 01 Mar 2022

Drug development is a long and arduous process, with much effort expended, often for little reward. It is imperative to optimise each step of the process and create solutions that maximise the chances of success for the patient. During the tissue analysis workflow, the Quality Control (QC) assessment is crucial to the production of high quality data for a robust, quantitative and streamlined analysis. Today, this step is manually performed and qualitative, with no simple and reliable solution to identify which tissue samples can be processed for further investigation.

In this article, David Wallis and Thomas Le Meur, two Data Science experts at Keen Eye, explain how Keen Eye's proprietary deep learning technology can significantly improve this process.

 

 

Digital pathology is an image-based discipline, combining the acquisition, management and interpretation of histopathological images. Whole slide imaging (WSI), along with the digitisation of entire pathology slides, has sped up the adoption of this discipline and revolutionised image analysis workflows, accelerating clinical routines and biopharma services. Technical (staining and scanning systems) and regulatory (FDA’s approval for WSI for primary diagnosis [1]) advances have helped make WSI possible, thereby increasing the availability of large image datasets. The ever-increasing numbers of digital images, coupled with the adoption of digital pathology platforms, has led to a strong need for the deployment of AI tools capable of (i) efficiently analysing large amounts of complex data and (ii) integrating within existing digital pathology image analysis workflows.

 

 

Tissue quality assessment is a crucial step in the digital image analysis workflow.

 

Each digital image analysis workflow starts with tissue processing: samples are collected, fixed and mounted on slides. They are then stained with hematoxylin and eosin (H&E) and scanned, before going through the QC test.

 

QC is a crucial step in the workflow, ensuring sufficient high quality data is available prior to biomarker evaluation. Currently, it is manually performed by a pathologist in a qualitative manner.

Most histology-based analyses will undergo QC, regardless of the pathology under investigation. It can typically be broken down into three main stages:

 

 

  • Image quality assessment, where image blur, artefacts, tears and folds are checked for.

  • Staining quality assessment, to look for the presence of staining oversaturation and noise.

  • Tissue assessment, where viable quality tumour area, amount of necrosis, etc., are quantified.

Following QC, a decision is made as to whether the sample quality is fit for downstream biomarker analysis or if a new biopsy is required.  It is therefore crucial to have a reliable, efficient and cost effective tissue quality assessment methodology in place beforehand.

 

 

A deep learning solution for tissue quality assessment.

 

Given the importance of accurate QC and its ubiquity in pathology workflows, we used our deep learning expertise to design a new solution for objective and quantitative tissue assessment. In oncology, one of the top reasons for failing QC is the lack of malignant cells [2].Keen Eye’s model has been specifically trained and optimised to identify  areas containing these cells. It takes as input H&E images and directly identifies and segments different tissue types (e.g. tumour cells, stroma, necrosis). It can then evaluate the segmentation to count the area of each tissue type and locate regions of interest. This can be used to decide whether the image passes the QC and if the corresponding tissue is a good candidate for further characterization. The results of the analysis can be directly seen and reviewed on Keen Eye’s digital pathology platform, allowing pathologists to quickly focus on any areas of interest highlighted by the analysis.

 

In summary, our solution:

 

  • Saves pathologist’s time. The application can run on multiple H&E slides in parallel and does not require any input from the pathologist to be launched.

  • Reduces workload. The increased automation helps pathologists focus on high-value tasks. The AI solution will also eliminate potential further complementary analysis.

  • Increases standardisation & reproducibility. The application eliminates human biases, meaning QC will be consistent across hospitals and centres. 

  • Extracts enriched data. The analysis is performed at a cellular level and extracts detailed quantitative information, compared to the mostly qualitative analysis of a pathologist.

 

Keen Eye’s deep learning model directly segments and counts the amount of different tissues in H&E slides, allowing the pathologist to quickly pass or fail the slides. This new solution can be simply implemented in existing digital pathology workflows, opening new avenues for optimised drug development and better patient stratification.

 

 

 

About the authors:

David Wallis
David Wallis
Data Scientist

David Wallis has a PhD in machine and deep learning applied to medical imaging from the Université Paris-Saclay (Paris, France). He is an enthusiastic Data Scientist who enjoys working within the collaborative team environment at Keen Eye. David strives to build powerful and smart models that can adapt to data from different sources.

Thomas Le Meur
Thomas Le Meur
Data Scientist

Thomas Le Meur is an engineer from ESIEE PARIS with a sound background in web development. His passion for data science led him to graduate at the University of California, Irvine, before joining Keen Eye. Thomas is particularly sensitive to the needs of the pathologists and is passionate about making a contribution to the revolutionary digitization of healthcare.

Find out how Keen Eye can support your digital image analysis:

 

References:

 

[1] Andrew J. Evans, Thomas W. Bauer, Marilyn M. Bui, Toby C. Cornish, Helena Duncan, Eric F. Glassy, Jason Hipp, Robert S. McGee, Doug Murphy, Charles Myers, Dennis G. O'Neill, Anil V. Parwani, B. Alan Rampy, Mohamed E. Salama, Liron Pantanowitz. US Food and Drug Administration Approval of Whole Slide Imaging for Primary Diagnosis: A Key Milestone Is Reached and New Questions Are Raised. Arch Pathol Lab Med. 2018. 

 

[2] Lazcano Rossana, Rojas Frank, Laberiano Caddie, Hernandez Sharia, Parra Edwin Roger. Pathology Quality Control for Multiplex Immunofluorescence and Image Analysis Assessment in Longitudinal Studies. Frontiers in Molecular Biosciences. 2021.