RAD-U-Net: Residual Attention-Based Dual-Path U-Net for Segmentation of Multiple Myeloma Cells in Microscopic Images for Clinical Diagnosis

Multiple myeloma, a malignancy originating in the bone marrow, poses significant health risks and can lead to severe systemic dysfunction and potentially fatal outcomes. Manual microscopic analysis of multiple myeloma cells remains a cornerstone in the diagnosis of this disease. However, this process is labor-intensive, time-consuming, and prone to human error. To address these challenges, we propose a novel approach leveraging advanced image processing and deep learning techniques for automated segmentation of multiple myeloma cells from microscopic images. We build a novel Residual Attention-based Dual-path U-Net architecture tailored specifically for this task. Our approach leverages deep learning techniques, including dual pathways for feature extraction with residual and attention mechanisms, to precisely detect multiple myeloma cells from complex microscopic images. Through rigorous experimentation, our model achieves outstanding segmentation performance, with a mean Intersection over Union (mIoU) of 91.50%, a dice score of 95.44%, mean average precision (mAP) of 96.46%. Despite its lightweight design, with only 5.7 million parameters, our model demonstrates remarkable accuracy as mentioned above, showcasing its potential for real-world clinical applications. Additionally, our proposed architecture maintains a balance between segmentation accuracy and computational efficiency because of reduced parameter, making it a more practical and effective solution for cell segmentation tasks, especially in scenarios where computational resources are limited or efficiency is a priority. However, there remains potential for further improvement in certain metrics, such as the mIoU and mAP, suggesting opportunities for refining the model to achieve even better segmentation accuracy.