Artificial Intelligence (AI) along with Machine Learning (ML) techniques has long promised to accelerate Digital Pathology (DP) based cancer diagnosis. Despite the consensus regarding the value of AI, the lack of visibility of how ML algorithms work, prevents their wider adoption for human in vitro diagnostic (IVD) in a highly regulated environment. A common ground becomes necessary in order to fully benefit from ML capabilities.
HalioDx Immunoscore® was the first immune scoring test validated for IVD use leveraging advanced image analysis. In brief, for each tumor sample, 2 slides are stained using an automated immunohistochemistry instrument: one with CD3 and one with CD8 ready-to-use monoclonal antibodies (HalioDx) followed by detection with DAB and counterstaining. Digital images of stained slides are obtained using a whole slide scanner and analyzed by a software program (Immunoscore® Analyzer, HalioDx)1.
Current workflow relies only on Computer Vision (CV) techniques for image analysis leading to the calculation of the Immunoscore®. We have used ML to improve HalioDx Immunoscore® software program, streamline the workflow, decrease hands-on and computation times.
In summary, to design the new workflow, each DP steps were considered as independent applications. CV remains applied to the cell detection. A Convolutional Neural Network, along with a UNET architecture, were used to recognize Regions of Interest (ROI) and image-related artifacts during the analysis. Intermediary validation steps by a trained operator were maintained in order to review CV and AI steps and guarantee a complete equivalence versus the standardized original DP protocol.
The Intersection over the Union of two regions (IoU) was used as performance and equivalency metric. Compared to Ground Truth, the ML algorithm improves the accuracy of the ROI detection versus the CV based algorithm, resulting in a dramatic decrease of the ROI computing time (from 3h to 5min) as well as in a reduced need for manual correction.
We demonstrated that ML applied to the Immunoscore® DP workflow for ROI detection results in reduced time-to result and overall improved robustness of the analysis. The equivalency study showed the importance of a well-curated dataset to maximize model's accuracy and performance. Finally, the verification and validation phase demonstrated the ML based workflow readiness for regulatory approval.
1Hermitte F. J Immunother Cancer. 2016 Sep 20;4:57. doi: 10.1186/s40425-016-0161-x.
Citation Format: Assil Benchaaben, Felipe Machado Guimaraes, Emmanuel Prestat, Alboukadel Kassambara, Mounia Filahi, Caroline Laugé, Thomas Sbarrato, Jacques Fieschi. Immunoscore® workflow enhanced by artificial intelligence [abstract]. In: Proceedings of the Annual Meeting of the American Association for Cancer Research 2020; 2020 Apr 27-28 and Jun 22-24. Philadelphia (PA): AACR; Cancer Res 2020;80(16 Suppl):Abstract nr 870.