Table 1 A comparison of the relative strengths and weaknesses of human pathology and computational histologic image analysis. The following is intended to illustrate the rational basis for the current project (summarized in Fig. 1). Considerations 1–2 favor the application of computational pathology as a screening tool for hypothesis generation. Considerations 4–6 outline some of the reasons for why human pathologists are arguably more suited for current clinical practice. Consideration 3 is a relative advantage of computer-based image analysis, but the concordance between and within observers for human-based histologic biomarkers vary widely. In the case of the SI score, the intra- and inter-observer concordance was found to be high, at least on tissue microarrays (see Results and Supplementary Figure 1).

From: Computationally-Guided Development of a Stromal Inflammation Histologic Biomarker in Lung Squamous Cell Carcinoma

  Computer-based histologic image analysis Human-based histology
1. Number of features assessed at one time Strength: Can be a relatively large number (e.g. > 1000) Weakness: Usually a relatively small number (e.g. < 10)
2. Bias in selecting features Strength: No inherent bias in favor of either neoplastic or non-neoplastic tissue Weakness: Potential bias towards epithelial biomarkers, since the classification of neoplastic tissue is a traditional focus of oncological pathology
3. Intra- and inter-observer variability Strength: The same algorithm analyzing the same digital image should give the same result every time Weakness: Reproducibility can be an issue for human pathologists, but this depends greatly on the feature assessed
4. Use in routine clinical practice Weakness: Not currently Strength: Human-based histopathology is the current gold-standard in clinical practice
5. Versatility across different clinical and pathologic settings Weakness: Algorithms developed for a specific application may not work well in other settings, e.g. an algorithm trained only on breast cancer examples may not interpret foci of adjacent benign breast lobules appropriately, if the system has not been trained to identify normal breast tissue. Strength: Human pathologists are currently more versatile than computer algorithms trained for specific applications. This represents a distinct advantage in the practice of general surgical pathology, which depends heavily on a very wide breadth of knowledge of different tissue types and pathological processes.
6. Availability of method Weakness: Currently not widely available; requires specialized software and hardware Strength: Currently widely available; can be used by people in any research or clinical laboratory