Table 2.
A proposed categorization of applications for spectroscopic imaging in histopathology
| Category of applications | Characteristics | |||
|---|---|---|---|---|
| Goal(s) | Typical design | Data characteristics | Applications | |
| Category I: Identification of tissue composition | Identify cells and constituents (i.e., pixel-level features) Identify cell types in tissue from compositional differences Measure biochemical composition or heterogeneity Detect extracellular components |
Many spectral instances (pixels) available within a single sample | Spectral differences between cells likely larger than diversity in populations Large spectral differences are expected and relatively simple Analyses are often univariate and are effective |
Well-established approach Many examples available (e.g., histology on many tissue types, microcalcifications, detection of disseminated tumor in lymph nodes) Output may serve as input for downstream analyses (e.g., with laser capture microdissection) |
| Category II: Identification of disease, physiology, or development | Identify disease, differences in physiology, or function in a sample (i.e., regional features) | Spectral instances include both pixels and whole regions (e.g., tumors in ducts) Many examples and patients available |
Subtle chemical differences Often affected by sample preparation Requires more sophisticated algorithms and careful spectral processing |
Detection of disease and heterogeneity within disease Detection of microenvironmental changes associated with disease progression Prediction of molecular expression Validation with IHC |
| Category III: Disease characterization | Determine severity of disease or subclasses that are clinically relevant (i.e., sample-level features) | Spectral instances and heterogeneity of tissue may need to be considered Determination at the patient level Examples and patients may be available for common diseases but heterogeneity measures and new classes may require prospective studies |
Subtle chemical differences often need to be considered in a spatial context May require multimodal information Requires significantly sophisticated algorithms and larger validation effort Unambiguous ground truth (e.g., disease grade) may be difficult to obtain |
Recapitulation of disease grades Discovery of subclasses of disease |
| Category IV: Prognostic and predictive, individualized analyses | Prognostication (outcome, regardless of therapy) and prediction (effect of a therapeutic intervention), with an ultimate goal of individualized results (i.e., human-level features) | Complete IR information and clinical and other information are incorporated into models for patient-level predictions New information from algorithms may be used to guide searches for biological causes of observed predictions Need to relate output of algorithms to images |
Multiple tissue components and spectral changes considered in a spatial context Often requires clinical and multimodal information Requires significantly sophisticated algorithms Retrospective validation combined with prospective validation is highly desirable |
Prognostication and prediction More individualized (group) and precise (patient-level) information that is often clinically actionable |
Abbreviations: IHC, immunohistochemical; IR, infrared.