Related Experiment Video
Updated: Aug 5, 2026
Predictive Immune Modeling of Solid Tumors
Published on: February 25, 2020
Real-World Evaluation of Artificial Intelligence (AI)-Based Tumor Cell Content Quantification for Molecular Tumor
Mariam Gachechiladze1, Jan-Niklas Runge1, Tobias Kull1
1Department of Pathology and Molecular Pathology, University Hospital Zurich, University of Zurich, Zurich, Switzerland.
Purpose:
Repeated evidence demonstrates limited reproducibility and accuracy of visual quantification (VQ) of tumor cell content (TCC) by clinical pathologists for downstream molecular testing. Artificial intelligence (AI)-based digital quantification (DQ) of TCC represents a promising alternative, yet real-world evidence from routine molecular diagnostic workflows remains limited. In this study, we evaluated the analytical performance and practical aspects of the analytical validation process of an AI-based DQ tool in routine molecular diagnostics.
Materials And Methods:
The clinically suitable AIM-TumorCellularity (PathAI) workflow was tested in molecular diagnostics for samples analyzed using comprehensive genomic profiling (FoundationOneCDx; Foundation Medicine Inc). The cohort included 300 nonpaired resection, biopsy, and cytology/cell block specimens from primary and metastatic breast (n = 66), lung (n = 117), colorectal (n = 40), pancreatic (n = 38), and prostate (n = 39) cancers, reflecting real-world diagnostic sample heterogeneity of a tertiary care center. We compared TCC estimates generated by pathologists' VQ, AI-based DQ, and molecular quantification (MQ) by bioinformatic deconvolution.
Results:
Agreement was lowest between VQ and DQ (Spearman rs = 0.38) and between VQ and MQ (rs = 0.44), while DQ showed stronger concordance with MQ (rs = 0.63). Single-cell validation against expert ground truth demonstrated high performance of DQ in tumor cell detection, with sensitivity of 0.985, specificity of 0.99, and accuracy of 0.99, based on 27,958 annotated cells across 60 regions of interest comparable with microscopic high-power fields. Analysis of preanalytical and analytical factors identified specimen type and cautery/crush artifacts as the main preanalytical contributors to DQ-VQ discrepancies, while overall variations in specimen cellularity were the dominant analytical factor.
Conclusions:
In summary, to our knowledge, this study provides the first comprehensive real-world evaluation of AI-based TCC quantification in routine molecular pathology workflows, highlighting its robustness, accuracy, the critical role of preanalytical standardization, and the need for pathologists' oversight in successful clinical implementation.
