TL;DR
This paper introduces a novel deep learning approach for label-free classification of white blood cells (WBCs), specifically distinguishing between activated and inactivated neutrophils. By leveraging a pre-trained ResNet-50 model and a custom microfluidic separation technique, the method achieves over 90% accuracy, offering a promising, less invasive alternative to traditional methods for rapid immune state monitoring.
What Problem Does This Paper Address?
Accurately counting and classifying white blood cells (WBCs) is crucial for diagnosing various diseases and monitoring a patient's immune status. Current methods, like manual counting or automated analyzers, often rely on staining, which can alter the cells' physiological state and morphology, potentially leading to inaccurate assessments. Furthermore, manual methods are labor-intensive, and automated equipment can be expensive and lack portability, hindering their use in point-of-care applications. This paper aims to bridge this gap by developing a label-free, accurate, and potentially portable method for WBC classification, focusing on the critical distinction between activated and inactivated neutrophils, which are key players in immune responses.
Methods
The core of this research lies in a two-pronged approach:
- Label-Free WBC Separation: The authors employ a spiral microfluidic channel to efficiently separate WBCs from red blood cells (RBCs) in whole blood. This method is chosen for its high throughput, speed, low cost, and portability, while minimizing WBC activation.
- Deep Learning Classification: A pre-trained ResNet-50 convolutional neural network (CNN) is utilized for feature extraction and classification. The network undergoes fine-tuning on the generated label-free WBC images. To adapt the model to single-channel grayscale images, these were superimposed to create three-channel inputs. The authors specifically chose ResNet-50 over deeper models like ResNet-101 or Inception V3 due to concerns about overfitting with limited datasets and the complexity of label-free feature extraction.
- Neutrophil Activation Classification: Beyond basic WBC type classification, the paper details a method to distinguish activated from inactivated neutrophils. This is achieved by analyzing morphological features like aspect ratio and roundness of the cells after convolution filtering. Specific thresholds are defined to classify neutrophils as activated or inactivated.
Datasets
The study generated a new dataset of label-free WBC images. Peripheral blood samples from healthy donors were used. After WBC separation via microfluidics, cells were stained for fluorescent imaging to obtain ground truth labels for bright-field images. The dataset was then enhanced through rotation and flipping to balance the classes, resulting in approximately 10,000 images per WBC type (Monocyte, Lymphocyte, Granulocyte). For the activated/inactivated neutrophil classification, 1771 granulocyte cells were tested.
Key Insights and Ideas
- Label-Free is Feasible and Advantageous: The research strongly demonstrates that accurate WBC classification is achievable without staining, preserving the cells' natural physiological state and morphology. This is a significant step towards less invasive and more reliable diagnostic tools.
- ResNet-50's Suitability: The choice of ResNet-50, with its residual connections, proved effective for extracting features from label-free images, which are inherently more complex than stained ones. The network's ability to balance accuracy and computational cost was a key factor.
- Morphological Cues for Activation: The paper highlights that simple morphological features like aspect ratio and roundness, when analyzed correctly, can reliably distinguish between activated and inactivated neutrophils, a critical indicator of immune response.
- Microfluidics as a Enabling Technology: The integration of microfluidic technology for cell separation is crucial, providing a high-throughput, portable, and gentle method for preparing samples for analysis.
Key Results
- Overall WBC Classification Accuracy: The ResNet-50 model achieved an accuracy of approximately 90% on the testing set for classifying WBCs into three types (granulocyte, lymphocyte, monocyte).
- Neutrophil Activation Classification: The method successfully classified activated and inactivated neutrophils. Out of 1771 tested granulocyte cells, 365 were identified as activated and 1328 as inactivated.
- Performance Comparison: ResNet-50 outperformed Inception V3 and Resnet-101 in terms of F1-score and stability (lower confidence interval fluctuations), suggesting it's a well-suited architecture for this specific task.
- Speed: The label-free classification method, particularly the morphological analysis for neutrophil activation, demonstrated a fast testing speed, processing 1771 cell images in just 60 seconds.
Limitations and Future Work
- Monocyte Classification Accuracy: The paper acknowledges that the accuracy for monocyte classification was lower (84.3%) compared to granulocytes (95.6%) and lymphocytes (94%). This is attributed to potential dataset imbalance (fewer monocytes in blood) and less distinct morphological features.
- Cell Detection Completeness: A limitation noted is that not all cells were detected during the morphological analysis for activated/inactivated neutrophils, leading to a slightly smaller number of analyzed cells (1693) compared to the total tested (1771).
- Further Improvement: The authors suggest that while their deep learning approach is convenient and time-saving, further improvements in feature extraction and network architecture could lead to even higher accuracy, potentially rivaling or surpassing manual feature extraction methods. Future work will focus on enhancing the network for better label-free feature extraction.
Summary updated