P8 Khartchenko et al.

Micro-immunohistochemistry meets machine learning: towards standardization

Anna Fomitcheva Khartchenko, Aditya Kashyap, Nuri Murat Arar, Pushpak Pati, Maria Gabrani, Govind V. Kaigala. IBM Research – Zurich, Switzerland

Background: Immunohistochemical stains frequently yield ambiguous results, which may lead to an improper treatment. These variations often arise due to IHC assay characteristics, specifically Ab concentration (C) and incubation time (t).  Standardization and quantification of antibody binding is critical to avoid false staining. We present a methodology to improve the quality of tissue stains and to aid delineation between false positives and false negatives.

Methods: We created microscale gradients of incubation time (0-300s) using anti-HER2, colocalized on individual tissue sections (primary and metastatic breast tumor, n=36. These gradients reduce ambiguity arising due differences in protocol and tissue-to-tissue variability. A machine-learning algorithm then assesses the quality of staining, through texture and contrast features chosen by random forest analysis and provides a tissue-type-specific staining quality metric (0<SQM<1). The SQM validates suitability of a stain for diagnosis.

Results: SQM classified stains into ‘acceptable’/‘non-acceptable’ for diagnosis (83% accuracy) and HER2+/- classes (92.1%).  Operational parameters to enhance quality while addressing tradeoffs between C and t were obtained in form of a 3D-manifold (SQM-C-t), e.g. with C=11-15µg/mL and t=85-130s provided an SQM≈1 for primary tumors.

Conclusions: Micrometer-scale-IHC assisted with machine-learning algorithms for analysis provides a rapid and tissue-efficient method for parameter standardization while identifying ‘true’ stains.