How to define and implement diagnosis assistants in tissue-based diagnosis (surgical pathology): A survey

Abstract

Digital pathology has started to enter the field of tissue - based diagnosis. It offers several applications, especially assistance in routine surgical pathology (tissue - based diagnosis). Diagnosis assistants are programs that assist the routine diagnosis work of a pathologist. Herein we describe how to appropriate design suitable algorithms. Theory: Tissue - based diagnoses derives from a) image content information, b) clinical history, c) expertise of the pathologist, d) knowledge about the disease. It can be transferred to a statistical decision algorithm (neural network, discriminate analysis, factor analysis, ... ). Image content information: Analysis of image content information (ICI) can contribute to medical diagnosis at different levels. The level depends upon the underlying disease (diagnosis) and the derived potential treatment. Pre - analysis algorithms include a) image standardization (shading, magnification, grey value distribution), and evaluation of regions of interest (ROI). ICI is embedded in three coordinates (texture, object, structure). Analysis of objects and structure require external knowledge (cell, nerve, vessel, tree, man, ... ). Texture is solely pixel - based and independent from external knowledge [1,2]. Algorithms: Stereology, syntactic structure analysis and measurement of object features (area, circumference, moments, ... ) are useful tools in combination with external knowledge and appropriate image standardization. Structure and texture parameters require the definition of neighbourhood (Voronoi, O'Caliaghan). Texture features are based upon algorithms that mimic time series analysis and can contribute to ROI definition and to disease classification [1, 2]. Material: Crude diagnoses have been automatically evaluated by the same algorithm from large sets of histological images comprising different organs (colon, lung, pleura, stomach, thyroid (> 1,000 cases). The trials resulted in a reproducible and correct classification (90 - 98 %). Conclusions: The applied algorithms can be combined to construct efficient diagnosis assistants. They can be extended to assistants of more differentiated diagnoses (inclusion of specific stains, clinical history, etc ... ). They can serve to formulate a general theory of "image information".