Table 3.

RBC classification and anemia detection methods

Methods	No. of images (Stain)	Performance metric	Remarks	Ref.
CHT, Heywood circularity factor, ANN, moment invariants, inclusion-tree structure, BPNN, PCA, SVM	150–1000 samples	80–99% accuracy for normal & abnormal RBCs	Lacks robustness	[18, 69, 147, 149, 154]
Morphological properties, Naive Bayes, K-NN, SVM, Sobel edge	626	94.6–96% accuracy for normal and sickle cells	Consider unsupervised classifiers for more RBC patterns	[128, 134]
CHT, WT, NN, decision tree, SOM, SVM	30–45 (Giemsa)	97–100% accuracy for sickle and elliptocytosis	Geometrical shape signature is used for detection process	[74–76]
Recursive partitioning, form factor	3878 cells	85% for discocytes, 83% for abnormal cells and 81% for sickle cells	Form factor invariant to cell size and provides useful information on cell shape	[121, 158]
Hybrid neural network	200 normal and 200 abnormal cells	91% accuracy for sickle, horn and elliptocytes	Considered only convexity index feature	[100]
DL, SVM	105 normal and 250 abnormal	Normal—100%, achantocyte—100%, sickle cell—90%, teardrop—100% and elliptocyte—73% accuracy using SVM	SVM classifier outperformed DL	[23, 26]
Rolling ball background, shape features, Naive Bayes, Bayesian classifier	1500 (Leishman)	98.2% precision for microcytic, macrocytic, sickle, teardrop, elliptocyte	Decision from CBC test measures is semi-automatic operation	[106]
ANN	1000 blood samples	Less computational time	Used RBG values—from Hb, MCH and RBC count	[162]
CNN , ELM	64,000 blood cells	94.71% accuracy	Images from multiple sources are used	[130]
U-Net	300 (MGG) and (Leishman)	91% sensitivity and 98% specificity	Results are shown for a variety of smear and stain	[116]
Inception recurrent residual CNN	352 WBCs and 3737 RBCs	100% for WBC and 99.94% accuracy for RBC	Model require larger number of network parameters	[27]
CNN	3737 labeled Cells	90.6% accuracy for 10 RBC classes	Label distribution was not homogeneous	[71]