Abstract
The accurate detection and segmentation of leukocytes is a challenging task in medical image processing. White blood cell images obtained under a microscope are easily affected by impurities, morphological variety, and overlapping phenomena, leading to inaccurate edge segmentation. To address these issues, this paper proposes a leukocyte detection method based on Mask R-CNN and an attention-mechanism-driven multi-scale feature fusion. The architecture integrates a Convolutional Block Attention Module (CBAM) into the Feature Pyramid Network (FPN), creating a Channel Spatial Feature Pyramid Network (CSFPN). This structure learns the weights of important channel features and the representation of critical feature regions. Additionally, a Skip-FPN module is incorporated to fuse low-level detailed information through short connections. Experimental results on the Kaggle open-source dataset demonstrate a mAP of 98.25% and a mIoU of 89.3%, showing significant performance improvements over baseline models.
Introduction
Automated white blood cell (WBC) analysis is a cornerstone of modern diagnostics for conditions like leukemia and anemia. However, clinical microscopy presents several challenges for computer vision algorithms:
- Impurities: Staining artifacts and debris in blood smears can be mistaken for cellular features.
- Morphological Variety: Subtle differences exist between WBC classes (e.g., neutrophils vs. eosinophils).
- Adhesion and Overlap: Clumped cells make it difficult to identify precise boundaries.
- Blurry Edges: Deep neural networks often lose fine-grained spatial detail, leading to imprecise segmentation.
Methodology: The C-Mask Architecture
The proposed C-Mask model enhances the standard Mask R-CNN framework with two primary architectural innovations: the Channel Spatial Feature Pyramid Network (CSFPN) and the Skip-FPN module.
Channel Spatial Feature Pyramid Network (CSFPN)
The CSFPN integrates the Convolutional Block Attention Module (CBAM) into the feature pyramid. This allows the network to focus on relevant cellular features while suppressing background noise and staining artifacts.
The CBAM component applies two distinct types of attention to the feature maps:
- Channel Attention: Learns "what" features (e.g., specific textures or nuclear shapes) are most indicative of WBCs.
- Spatial Attention: Learns "where" the most relevant information is located, helping the model ignore debris.
Skip-FPN
Standard FPNs pass information from bottom to top, which can cause high-resolution details to be "washed out" in deeper layers. The Skip-FPN introduces shortcut connections from early, high-resolution layers (specifically conv2) directly to the upper levels of the pyramid. This preserves the low-level edge information necessary to accurately separate cells that are touching or overlapping.
Experimental Results
The study utilized a subset of 4,000 images from the Kaggle WBC dataset, balanced across four leukocyte types: Eosinophils, Lymphocytes, Monocytes, and Neutrophils. The C-Mask model demonstrated superior performance compared to baseline architectures:
- Detection Accuracy: Achieved a mAP of 98.25%, a 1.25% improvement over standard Mask R-CNN.
- Segmentation Quality: Reached a mIoU of 89.30%.
- Robustness: The False Positive Rate (FPR) dropped to 1.75% (compared to 3.00% for the baseline), proving the effectiveness of the attention mechanism in ignoring debris.
In comparative tests, C-Mask significantly outperformed U-Net (86.8% mAP) and FCN (72.1% mAP). Qualitative results showed that C-Mask was particularly effective at identifying the distinct multi-lobed nuclei of neutrophils and eosinophils, which are often difficult to segment precisely.
Conclusion
By combining attention mechanisms with multi-scale feature fusion, the C-Mask architecture effectively addresses the core difficulties of leukocyte image analysis. Future work will focus on expanding the dataset to include rarer cell types and bone marrow samples, as well as optimizing the model for real-time clinical use.
Summary updated