Using Computer Vision to Identify and Count Blood Cells helps in medical diagnosis

How can computer vision models identify and count CBCs?

The workflow for developing a computer vision (CV) model to accurately identify and count CBCs is shown in Figure 1.

Dataset gathering

To train the YOLO model, a public dataset of annotated blood cell images known as the Blood Cell Count Dataset (BCCD) is used by the authors. The BCCD consists of 364 annotated smear images, divided into the train (300) and test (60) subgroups. The validation set is randomly chosen from the training images with annotations (60). The modified dataset can be downloaded here. Another dataset including 100 images, with a resolution of 3246 × 2448 pixels captured with a Nikon V1 camera mounted on a Nikon ECLIPSE 50i microscope with a magnification of 100x, is used for the comparison purpose.

YOLO algorithms approach object detection like a regression problem. They can rapidly predict both image classes and locations using only one forward propagation pass through the network. Images with 448 × 448 pixels are split into 7 × 7pixel grid cells and each predicts two bounding boxes and confidence scores for the boxes. The network architecture includes 24 convolutional layers and 2 fully connected layers using GoogLeNet. The authors used the fastest algorithm, Tiny YOLO, which uses 9 instead of 24 convolutional layers.

Dataset training

The Tiny YOLO model is trained for 20 different classes. To achieve higher accuracy for blood cell identification, the model is optimized for three additional classes, WBC, RBC, and platelets. For each bounding box, five values along with class probabilities are predicted. These values include the probability for the presence of an object in a grid cell, the x and y coordinates of the object, and the height and width of the object.

The authors obtain loss and moving average loss in each step of the training process for a total of 4500 steps using two different learning rates of 10^-5 and 10^-7 for steps 1–2500, and 2501–4500, respectively. A lower learning rate in the latter steps results in a better convergence. The learning curve of the YOLO framework for blood cell identification is depicted in Figure 2. The minimum moving average loss occurred on step 3750 at the learning rate of 10⁻⁷.

Blood cell identification and counting method

The threshold value is calculated by averaging the absolute error between ground truths and the model estimation at different threshold values. The threshold for each class equals a minimum average absolute error in the validation dataset. The proposed model occasionally double counts platelets, which is solved by using K-nearest neighbor (KNN) and intersection over union (IOU) in each platelet.

For counting cells, the authors employ class labels. In fact, the trained model returns three kinds of labels: ‘RBC’, ‘WBC’, and ‘Platelets’. The total number of ‘RBC’ labels shows the number of RBCs in a smear image. Figure 3 displays a platelet detection example in which a platelet is detected twice by the YOLO algorithm. This issue is tackled using the proposed KNN and IOU-based technique.