Table 1 Overview of the different diagnostic tests, the current challenges and known confunders for the clinical implementation of AI-based methods, and the requirements for a successful implementation.
From: How artificial intelligence might disrupt diagnostics in hematology in the near future
Cytomorphology | Cytogenetics | Immunophenotyping | Molecular genetics | |
|---|---|---|---|---|
Method | Microscopy | Chromosome banding analysis | Multiparameter flow cytometry | Genomic analysis |
Aim | Identification and characterization of cell populations based on morphology | Identification of cytogenetic abnormalities | Identification and characterization of cell populations based on light-scattering properties and antigen expression patterns | Identification of individual molecular profiles |
Challenges | Differentiation between artefacts and informative material (=cells) | Identification and selection of individual chromosomes | Accurate representation and transformation of the raw data | Data matrix usually sparse and informative signals might be lost in noise |
Correct identification of borders (very dense regions with overlapping cells) | Correct identification of structural abnormalities | Meaningful combination of various data types and differentiation between absence of information (e.g., insufficient coverage) and true negative results | ||
Extraction of features that allow the differentiation of maturation states of the same cell type | Lack of knowledge for annotation and interpretation of variants in coding and especially non-coding regions | |||
Confunders | Resolution, image capturing, image cropping are not standardized between laboratories | Different banding and staining methods | Different methods for data pre-processing and data/image transformation | Plethora of methods for the identification of features (CNV, SV, SNV, Fusions, etc), data transformation, and dimensionality reduction with limited concordance and individual biases |
Unbalanced data sets for training with rare cell types being underrepresented | Unbalanced data sets for training with rare structural abnormalities being underrepresented | |||
Requirements for final implementation | Time and cost efficient digitization of glass slides | Harmonization and standardization of used antibodies | Harmonization of gene panels | |
Standardized, automated systems for the recording of digital microscopic images | Standardization of analysis pipelines and variant interpretation | |||
Balanced training data capturing as much biological and technical variety as possible | ||||
