Abstract:
It is well known that the current methods used for blood cell segmentation and counting
are very tedious and are highly prone to different sources of errors especially if the
density of cells is very high, and overlapping of cells are more common. In addition,
those techniques did not provide full information related to blood cells like their shape
and size, which play important roles in clinical investigation of serious blood related
diseases such as leukemia and sickle cell disease.
Among currently emerging techniques for blood cell detection and counting, image
processing and neural-network-based methods are mostly used to automate the tasks
and provide accurate diagnostic results. This research used YOLOv2 model, which is
currently the state of the art for real time object detection and classification. The model
is faster and more accurate than any other currently available real time object detection
networks such as CNN, RCNN, fast-RCNN due to its unique features consisting of
global vision, optimization and others. It is used to automatically detect classify, count
and measure the morphological parameters of blood cells.
The model was trained on 1558 images and 2733-labeled blood cells with network a
resolution of 608x608, different batch sizes (16, 32, 64 and 128), a learning rate for
more than 15000 number of steps in which the loss converged from 234.5 to 1.54. Nonmaximum suppression was applied on the output of the network to remove duplicated
detections of objects based on class label, score, overlapping area and non-maximum
suppression threshold and finally it was tested on 26 images containing 1454 red blood
cells, 159 platelets, 3 basophils, 10 eosinophils, 24 lymphocytes, 13 monocytes and 28
neutrophils. The network achieved detection and segmentation of blood cells with
average accuracy of 80.6% and precision of 88.4%.
After classification and counting, measurement of area, diameter and aspect ratio of
cells were calculated using image processing and the results were evaluated using a
manual method (ImageJ), resulting in a mean accuracy of 92.96%, 91.96%, 88.736%
and 92.7% in the measurement of area, aspect ratio, diameter and counting of cells
respectively
