Immunofixation electrophoresis image interpretation using transfer learning method

Area of Science:

• Medical Diagnostics
• Artificial Intelligence
• Computational Pathology

Background:

• Immunofixation electrophoresis (IFE) is the standard for diagnosing plasma cell disorders but is complex and subjective.
• Inter-observer variability and time constraints challenge current IFE interpretation.
• Developing automated interpretation methods is crucial for efficient and accurate diagnostics.

Purpose of the Study:

• To develop and evaluate deep learning models for automated IFE image interpretation.
• To reduce subjectivity and improve the speed of monoclonal protein identification.
• To assess the performance of different deep learning architectures for IFE analysis.

Main Methods:

• A dataset of 5226 IFE images was utilized, split into training (80%) and testing (20%) sets.
• Two deep learning approaches were developed: a two-stage (binary then subclass) and a single-step multi-class model.
• Transfer learning with a YOLOv11 architecture was employed, with performance metrics including accuracy, precision, recall, and F1-score.

Main Results:

• The two-stage model achieved 94.76% accuracy for binary classification and 91.28% for subclass identification.
• The single-step multi-class model showed a slightly higher overall accuracy of 92.12%.
• Both models demonstrated high accuracy for dominant patterns but struggled with underrepresented or visually similar classes.

Conclusions:

• Deep learning models can accurately classify IFE images, demonstrating potential for laboratory workflow integration.
• The two-stage approach offers better interpretability, while the multi-class model provides scalability and efficiency.
• Future work should focus on expanding datasets for rare patterns and improving classification of low-frequency classes.