Deep Learning for Microscopic Image Segmentation in Multiple Myeloma: A Comparative Study

Multiple myeloma (MM) is a malignant disorder of plasma cells that presents considerable diagnostic challenges due to cellular overlap and morphological complexity in bone marrow aspirate images. Traditional diagnostic approaches rely on manual examination, which is very labor-intensive and prone to variability. While deep learning has shown promise for automated cell segmentation, standardized evaluation across models remains limited. This paper introduces a unified benchmarking approach that applies a cell-centric segmentation preprocessing workflow to evaluate the performance of three state-of-the-art deep learning models for plasma cell segmentation including U-Net, SAM, and Med-SAM. Using the SegPC2021 dataset, individual cell patches are extracted standardized training and evaluation procedures are applied across all models. The results show that the fine-tuned SAM model achieves the highest segmentation performance, with a Dice score of 95.1% and mIoU of 80.6%.