Postdoctoral fellow of Faculty of Mathematics and Informatics (MIF) of Vilnius University (VU) dr. Dovilė Žilėnaitė-Petrulaitienė won the position for postdoctoral fellowship “Digital image data-driven analysis for tumor microenvironment assessment in breast carcinoma”. How important is this win? Dr. D. Žilėnaitė-Petrulaitienė will search for new biomarkers in pathological images for early diagnosis of breast cancer.
The research will last a few years – from 2022 to 2024. “The search for informative biomarkers for early cancer diagnosis has been developed using heuristic methods for a long time. With the advent of high-level information extraction capabilities using deep learning technologies, this can be successfully transferred to medical image analysis, automating and accelerating biomarker searches. Therefore, this research has a very important practical relevance, which will be able to be used in practice”, says the fellowship supervisor, assoc. professor dr. Linas Petkevičius.
The new era of digital pathology based on machine learning in clinical practice
For many years, the assessment detection of morphological features such as tumor size, grade, lymph node involvement, and histologic type is the main and sometimes the only pathological method to determine the patient’s diagnosis, prognosis, and response to treatment.
However, the advent of artificial intelligence and machine learning technologies and greater attention to precision medicine paves the way for the era of digital pathology in clinical practice, allowing not only more accurate quantitative and spatial assessment of biomarker expression than the visual assessment of a pathologist but also allows to extract information invisible to the eye. Today, various tools based on artificial intelligence significantly contribute to the complex encryption of pathophysiology, allowing the discovery of new biomarkers and drug targets.
In 2015, dr. Dovilė Žilėnaitė-Petrulaitienė joined the team of the National Center of Pathology together with prof. Arvydas Laurinavičius and his team of scientists. They started complex experimental work creating an analytic methodology to analyze local profiles of cancer cells and microenvironment properties, generating multiparametric and spatial data from digital microscopy images.
Using deep learning and artificial neural network models for tumor tissue classification, applying spatial statistical methods and clear mathematical rules, an exceptional methodology for tumor tissue research was created. By applying this method, it is possible to automatically determine the tumor-stroma interface zone, calculate indicators for anti-tumor response assessment, and most importantly, predict the overall survival of patients.
In dr. D. Žilėnaitė-Petrulaitienė’s doctoral dissertation (2021) ” Assessing Immunohistochemistry Biomarkers in the Spatial Context of the Microenvironment of Hormone Receptor-Positive Ductal Breast Carcinoma by Digital Image Analysis” using this high-efficiency, automated and repeatable data set-based method, it was demonstrated that in whole slide histological tumor images, it is possible to detect breast cancer tissue microenvironment components, extract new and clinically useful immune response parameters and predict a higher probability of survival by analyzing the indicators describing the spatial characteristics of the immune infiltrate in tumor samples.
Next steps: to combine features for direct biomarker assessment
This research has shown that prognostic modeling can be obtained based entirely on the digital image analysis of histological images and reflected three biological and prognostically independent features of breast cancer. These three indicators surpassed conventional clinical and pathological parameters. It should be also noted, that in this work, it was also detected that the new biomarker – CD8+SATB1+ T cells, provided stronger prognostic information than CD8+ lymphocytes assessment and reflected the activated state of immune cells.
However, these are only preliminary data – the obtained models require further adjustment and validation in a larger patient cohort. On the other hand, this work did not consider the architectural properties of the tumor microenvironment – the arrangement of tissue collagen fibers, which can also significantly contribute to tumorigenesis and help determine the prognosis of patients. In order to complete this research, more comprehensive computational methods and statistical modeling are required to accomplish this work. We believe that the combination of explicit, image-based computational biomarkers and implicit, artificial intelligence-derived features will produce the best set of methods to support clinical decisions in precision medicine.
The postdoctoral fellowship is financed by the Research Council of Lithuania (LMTLT; agreement No.: S-PD-22-86).