Benefits of Automated Slide Reading as a Way to Decrease Human Error in Clinical Trials and thus in Patient Care

Sep 14, 2017 5:28:20 PM

Human error is natural. It is a result on how the human brain works and its biological limitations. Though human error is inevitable and also normal in highly specialized experts such as pathologists, it should not end in failure and affecting the quality of diagnosis and the selection of the adequate treatment. However, mistakes also help us and any system to learn. The error rate for complex logic errors (which applies to reading and interpreting tissue slides) is about 5% overall. Despite extensive training and education, it does not seem to drop below this natural threshold.

Development of image analysis applications by applying Machine Learning and Artificial Intelligence methods shows many advantages for the image analysis experts as well as for the pathologist. Machine Learning algorithms learn to recognize images and patterns almost the same way as humans do but not necessarily dependent on handcrafted features like color intensities and object morphologies. Adding Machine Learning methods into image analysis workflows for diagnosis purposes add a number of benefits to such workflows. There is no need to formalize complex “handcrafted features” (e.g. cancer regions versus the surrounding stroma) the pathologist just needs to annotate. The system can be trained on a variety of samples (possibly large data sets) to achieve a robust and sustainable recognition solution and new data samples can easily be added into the trained object model in order to continuously increase accuracy even below a 5% error rate. 

In the immediate future, the need for high level requirements of highly sophisticated algorithms and standardized methods will become a possibility for the diagnostic environments. But this will require that the methods and algorithms are not only able to analyze big data but also to analyze very complex data sets to achieve a maximal exploitation of available information. The maximum extraction of such information from big data in tissue sections with quantitative methods such as Tissue Phenomics requires also an easy-to-access platform that can be used by experts from multiple domains, e.g. pathologists, biomarker experts and image analysis specialists to serve as a collaboration and interaction tool. Co-registered multi-viewing opportunities and the simultaneous analysis of multiple biomarkers on single or consecutive sections (“virtual multiplexing”) expand the value and reliability of the data generation and support the decision-making process. For these reasons a Tissue Phenomics user interface (UI) allows not only slide / image navigation, the changes of magnification, snap shots, contemporary viewing of several images (for example HE and IHC images synchronously), but also employs fully automated co-registration of serial sections and fully automated transfer of annotated regions (RoI) from the annotated single slide into the remaining co-registered serial slides.

Automated image analysis with an AI / machine learning backbone also enables the development of a “new generation” of diagnostic approaches by improving existing disease grading in cancer, based on the quantification of the morphological features in such tissue images. Any improvements are based on newly defined biomarkers, identified by unlocking the “new knowledge” and providing deeper insights into tissue based information. Those tissue based biomarkers are quantitative descriptors of functional, morphological and spatial patterns, which are hopefully correlated to the disease progression or response prediction. This approach fits well into the context of precision medicine, because of its ability to support other prognostic and predictive diagnosis and to help the physicians to better define the appropriate treatment for each individual patient. First explorative studies have already demonstrated that such new quantitative tissue patterns can be highly predictive and prognostic and can stratify patients much better than previously defined tissue based scoring solutions.

Digital pathology workflows provide tools and resources for pathologists to effectively and accurately read and interpret slides in a robust and reliable manner and share information via an integrated digital pathology platform to further analyze slides with image analysis algorithms like those for machine learning and visual recognition applications and to create and distribute reports for the exchange of documents with clinical colleagues.

Future digital pathology workflows will also include cloud integration to enable pathologist to start the process of annotations and analysis of images by simply uploading images to the appropriate web workspace, add markers and annotations requesting the analysis. If not available yet, experts can provide own novel image analysis solutions by developing the appropriate image analysis applications. Once the application is ready, it could be added to the pathologist’s web workspace on the server. Pathologists will apply and validate such solutions on existing and new images in a systemic way even independent of natural human error-prone inconsistencies to further optimize the quality of diagnosis and therapy in clinical trials and patient care.

Author: Dr. Ralf Huss, Chief Medical Officer, Definiens