Digital pathology opens new opportunities for artificial intelligence applications in disease diagnosis and treatment. Major benefits come with extraction and quantification of novel features of pathology, often not visible by usual microscopy examination. The field is rapidly developing in many areas of medicine.
We started digital pathology research in 2010 with image analysis-based quantification1, 2 and exploring the benefits of multivariate analytics of image analysis data to generate combined image-based biomarkers3, 4. Further, we performed accuracy and calibration experiments for image analysis-based quantification5, 6, 7, 8.
In 2015, we proposed a methodology based on hexagonal tiling of image analysis data to quantify intra-tumor heterogeneity of biomarker expression9; this allowed to generate rich data set to compute spatial indicators. We conceptualized this approach as “comprehensive immunohistochemistry”10. The heterogeneity indicators were demonstrated in later studies as independent prognostic factors, often exceeding the informative value of the average level of the biomarker expression11, 12.
In 2020, we published another hexagonal grid-based methodology to automatically detect tumor-host interface zone and compute immune cell density profile across the zone13; this actually measures the “willingness” of immune cells to enter the tumor (immunogradient) and provides independent prognostic value11, 13, 14. Compared to other methods proposed to measure immune response in the tumor microenvironment, the interface zone immunogradient provides quantitative directional assessment in the very frontline of tumor-host interaction.
We demonstrated that intratumoral heterogeneity- and immunogradient-based computational biomarkers could be combined to predict patient survival without requirement for any conventional clinical or pathology data12. Also, computational augmentation of a single CD8 immunohistochemistry image data generated three independent prognostic features15. Tumor-associated fibrillary collagen architecture, retrieved from 1 mm breast cancer tissue microarray core by deep learning network and assessed by multiple morphometry metrics, could predict survival of the patients16. The methods were recently reviewed in the context of artificial intelligence applications for tumor pathology17, 18