Development and validation of a deep learning-based protein electrophoresis classification algorithm.

Area of Science:

• Biomedical imaging analysis
• Artificial intelligence in diagnostics
• Clinical biochemistry

Background:

• Protein electrophoresis (PEP) is crucial for analyzing protein status in various diseases, including monoclonal components, inflammation, and antibody deficiency.
• Manual interpretation of PEP is labor-intensive and can suffer from inter-observer variability.
• Developing automated methods can improve efficiency and consistency in PEP analysis.

Purpose of the Study:

• To develop and evaluate a deep learning-based algorithm for classifying protein electrophoresis (PEP) images.
• To supplement manual PEP interpretation and improve inter-observer reliability.
• To provide an automated tool for screening PEP results.

Main Methods:

• A dataset of 2,578 gel and densitogram PEP images was collected and split into training, validation, and test sets.
• Image preprocessing included color-to-grayscale conversion and histogram equalization.
• The processed images were input into neural networks for classification of six major findings.

Main Results:

• The deep learning algorithm achieved high accuracy in classifying PEP images, with densitogram analysis showing area under the receiver operating characteristic curve (AUROC) from 0.873 to 0.989 and accuracy from 85.2% to 96.9%.
• Gel image analysis yielded AUROC values ranging from 0.763 to 0.965 and accuracy from 82.0% to 94.5%.
• The algorithm demonstrated robust performance across different diagnostic categories.

Conclusions:

• The developed deep learning algorithm shows strong performance in classifying protein electrophoresis images.
• This AI tool can serve as a valuable auxiliary for screening PEP results.
• The algorithm is particularly beneficial in healthcare settings with limited access to specialists.