Table 2.

RBC segmentation and counting methods

Methods	No. of images(Stain)	Accuracy(%)	Remarks	Ref.
K-means clustering, WT	60 (Giemsa )100 (Wright–Giemsa)	93–98.9	Robustness is not explained	[11, 140, 157]
Iterative structured circle detection, circlet transform	100	95.3	Incorrect hole filling leads to errors To improve initial RBCs mask for accurate segmentation	[28, 143]
Graph algorithm	98	99	Considered only non-overlapped cells	[133]
Parametric template matching, PCNN	900 cells	90–95.7	Require prior knowledge about the appearance of the cell	[16, 46, 98]
YOLO algorithm	364	96.1	Satisfactory performance	[20]
HSV conversion, morphological operations	200 (Giemsa)	96	Used uniform staining and illumination	[125]
Pixel relationship	10 (MGG)	83	Occluded objects are rejected before the later stages	[80]
Canny, LOG, Sobel	20–30	85–93	Normal RBCs Less samples	[43, 60]
K-curvature, circumference and ellipse adjustments	66	98	Images are not preprocessed to reduce execution time	[81]
Blob analysis, WT	10 blood samples	90–96	Need optimization to get accurate results	[57, 86]
CNN AlexNet	5772	90	Average execution time was 227 ms	[135]
Canny edge, MLP	59 RBCs and 59 non RBCs	74–88	Increase training images	[123] [19]
K-medoids, distance transform	1000 (Wright)	98	Processing of central pallor of RBCs consume more time	[15]
HT	500 subjects	91–94.9	Many tunable parameters	[72, 110, 148, 155]
Deep neural network models	100 (MGG)	Indices lie within the 10% of Sysmex reported value	Considered only normal blood smear images	[51]
CHT, NN, SVM	368	98	Achieved low false negative rate	[101]
LAB, YCbCr color space, CHT	108(Wright)	81–91	Computational time is more	[105] [161]
Region proposal	180 (Wright)	96-98	Tested on ALL-IDB and MP-IDB datasets	[65]
Semantic segmentation	108 (Wright)	91–97	More labeled images are required	[143, 150, 151]