DeepLIIF

DeepLIIF infers multiplex immunofluorescence (mpIF) channels from immunohistochemistry (IHC) and provides quantitative single-cell IHC scoring using a multitask deep learning framework.

Key Features:

• Multitask deep learning framework: Performs stain deconvolution/separation, cell segmentation, and IHC quantification in a unified model.
• Stain deconvolution/separation: Separates overlapping stains on IHC slides and reconstructs mpIF channels from IHC images.
• Cell segmentation and classification: Leverages a LAP2beta nuclear-envelope stain with over 95% cell coverage to improve cell delineation and segmentation accuracy.
• Quantitative single-cell IHC scoring: Produces single-cell protein expression measurements and generalizes from clean to noisy or artifact-laden IHC images.
• Broad marker compatibility: Trained and validated on diverse nuclear and non-nuclear markers including Ki67, CD3, CD8, BCL2, BCL6, MYC, MUM1, CD10, and TP53.
• Benchmark evaluation: Methodology evaluated against publicly available benchmark datasets and compared with pathologists' semi-quantitative scoring.

Scientific Applications:

• Diagnostic pathology: Supports quantitative biomarker expression reporting from routine IHC for clinical decision-making.
• Research studies: Enables cellular- and single-cell-level analyses by inferring mpIF from conventional IHC without additional mpIF staining.
• Biomarker validation and comparison: Facilitates objective comparison of marker expression across samples and against pathologist scoring.

Methodology:

Trains a multitask deep learning model on a dataset of co-registered IHC and mpIF staining on identical slides used as ground truth to learn translation of IHC into separated mpIF channels while performing cell segmentation and classification, with evaluation against publicly available benchmark datasets and pathologists' semi-quantitative scoring.