Immunofixation electrophoresis image interpretation using transfer learning method

Immunofixation electrophoresis (IFE) is the gold standard for identifying monoclonal proteins in plasma cell disorders, but its interpretation is complex, time-consuming, and prone to inter-observer variability. This study aimed to develop deep learning models that can rapidly and accurately interpret IFE images and reduce subjectivity in diagnostic evaluation.

A dataset of 5226 IFE images from University Hospital was labeled by experts and divided into training (80 %) and testing (20 %) sets. Two classification strategies were explored: a two-stage model (Approach-1) involving binary classification followed by subclass categorization, and a single-step multi-class model (Approach-2). A transfer learning method using a pre-trained YOLOv11 architecture was applied. Model performance was evaluated using accuracy, precision, recall, F1-score, and confusion matrices.

Approach-1 achieved 94.76 % accuracy in binary classification and 91.28 % in subclass identification. Approach-2 yielded slightly higher overall accuracy (92.12 %). Both models demonstrated high reliability in detecting dominant patterns (e.g., IgG-κ, IgA-λ), but lower performance in underrepresented or visually similar classes (e.g., IgM-λ, light chain-κ). The confusion matrix analyses demonstrated that both approaches achieved high accuracy in major patterns, while misclassifications occurred predominantly between visually similar classes, and lower performance was observed in less represented patterns.

The proposed deep learning models successfully classify IFE images with high accuracy and generalization, showing promise for integration into laboratory workflows. Approach-1 may be preferable for interpretability, while Approach-2 offers scalability and efficiency. Expanding the dataset to include rare patterns and investigating methods for improving classification in low-frequency classes are important for future development.