Abstract / Summary
Abstract Serum immunofixation electrophoresis (IFE) is the reference method for identifying and characterizing monoclonal immunoglobulins in plasma cell dyscrasias; however, its interpretation remains time-consuming, operator-dependent, and subject to inter-observer variability. This study evaluated a deep learning-based decision-support system (ImmunoGUI) for automated IFE interpretation intended to support the laboratory reporting workflow. A dual-stage framework was developed to isolate electrophoretic regions of interest and automatically generate standardized textual reports, paired with visual attention maps for diagnostic explainability. The system was evaluated on real-patient datasets across six canonical intact monoclonal immunoglobulin classes (IgGκ, IgGλ, IgAκ, IgAλ, IgMκ, and IgMλ). Models were trained using combinations of real-patient and synthetic IFE images and evaluated on independent real-patient test cohorts. The first-stage model achieved 89.5% overall binary accuracy, with 91% sensitivity and 88% specificity. The second-stage structured reporting model achieved a Top-1 diagnostic accuracy of 79.4%, reaching 95.5% Top-3 accuracy when evaluated as a decision-support hypothesis ranker. Visual attention maps reliably highlighted target monoclonal bands, providing intuitive spatial explainability to support clinical review. ImmunoGUI offers a practical, computationally lightweight decision-support tool to semi-automated IFE reporting. Although further validation on larger multicentre datasets will be required before routine clinical implementation, the system has the potential to help in reducing manual reporting workload and enhancing standardization in routine IFE screening under expert human oversight.