Multiple myeloma is a heterogeneous malignant disease characterized by clonal plasma cells expressing either κ or λ immunoglobulin light chains. The aim of this thesis is to investigate whether morphological characteristics of plasma cells differ according to the expressed light-chain type. We develop a deep learning approach based on a residual convolutional neural network (ResNet-18) to classify cells using multi-channel fluorescence microscopy images.
The results indicate that morphological differences between κ- and λ -positive cells are not stably represented in the image-based feature space under four-class and binary classification formulations. However, after reformulating the task as a three-class classification problem separating κ-positive cells, λ -positive cells, and non-clonal cells, an optimized reduced-capacity architecture achieves robust classification performance using three imaging channels, indicating biologically meaningful morphological differentiation independent of a direct clonal marker signal.
|
