A machine learning algorithm for simulating immunohistochemistry: development of SOX10 virtual IHC and evaluation on...

Area of Science:

• Computational pathology
• Digital pathology
• Machine learning in diagnostics

Background:

• Immunohistochemistry (IHC) is a crucial diagnostic tool in pathology.
• Current IHC methods are costly, time-consuming, and consume valuable tissue.
• Existing predictive models lack the spatial resolution for cell-specific analysis.

Purpose of the Study:

• To develop a machine learning algorithm capable of predicting individual cell immunophenotype using only hematoxylin and eosin (H&E) stained slides.
• To create a virtual IHC (vIHC) assay that reduces the need for physical IHC staining.
• To demonstrate the feasibility of predicting SOX10 nuclear staining from H&E images.

Main Methods:

• Digital whole slide images were acquired from H&E stained slides.
• Slides were destained and subsequently stained with SOX10 IHC.
• A convolutional neural network was trained on registered H&E and SOX10 IHC images to predict SOX10 positivity.
• Machine learning techniques were used for cell segmentation and annotation.

Main Results:

• The virtual IHC (vIHC) neural network achieved an area under the curve of 0.9422 in predicting SOX10 nuclear staining.
• The model successfully segmented millions of cells based on H&E and IHC data.
• Qualitative evaluation by a dermatopathologist indicated potential clinical utility.

Conclusions:

• This proof-of-concept study demonstrates the feasibility of neural network-driven virtual IHC assays.
• The vIHC approach shows promise for cost savings, time efficiency, and reduced tissue consumption in pathology.
• Further optimization is required for widespread clinical adoption and improved accuracy.