DeepHistoClass

DeepHistoClass classifies immunohistochemistry (IHC) images using deep learning to produce confident multilabel cell-type and protein annotations for integration with single-cell transcriptomics and the Human Protein Atlas.

Key Features:

• Multilabel Classification: Performs multilabel classification of complex IHC images and was demonstrated on 7848 human testis images corresponding to 2794 unique proteins from the Human Protein Atlas (HPA).
• Hybrid Bayesian Neural Network: Employs a Hybrid Bayesian Neural Network trained and tested using manual annotations from eight distinct cell types within the IHC images.
• DeepHistoClass Confidence Score: Implements the DeepHistoClass (DHC) Confidence Score as an uncertainty metric to identify reliably classified images and potential manual annotation errors.
• Improved Diagnostic Performance: The DHC Confidence Score increases average diagnostic accuracy from 86.9% to 96.3%.

Scientific Applications:

• Digital Pathology: Enables multilabel classification of IHC images for digital pathology analyses across tissue types.
• Protein Mapping Initiatives: Facilitates large-scale protein mapping projects such as the HPA by linking IHC-based protein localization with transcriptomics datasets.
• Error Identification: Detects potential manual annotation errors in IHC datasets via its uncertainty metric, improving annotation reliability.

Methodology:

Trains and tests a Hybrid Bayesian Neural Network on manually annotated IHC images and computes the DeepHistoClass Confidence Score as an uncertainty metric to assess classification confidence.