Cell Phenotype Classification Based on Joint of Texture Information and Multilayer Feature Extraction in DenseNet

Table 6

Comparison results of state-of-the-art methods on the Hep2 dataset in terms of accuracy, precision, and F1 score (%).

Approach	Type of features	Accuracy (%)	Precision (%)	F1 score (%)
VGG-16 [1]	Deep	86.36 ± 0.84	—	85.00
VGG-16 + transfer learning [1]	Deep	86.91 ± 0.50	—	90.00
Inspection-v3 [1]	Deep	82.37 ± 0.36	—	81.00
Inspection-v3 + transfer learning [1]	Deep	84.65 ± 0.42	—	89.00
Scale space + SVM [17]	Handcrafted	87.99	—	84.05
Ensemble inception-v3 and Resnet152 and inception-Resnet-v2 [20]	Deep	94.98 ± 1.13	—	—
Ensemble inception-v3 and Resnet152 [20]	Deep	94.78 ± 1.05	—	—
ResNet152 [20]	Deep	92.28 ± 1.59	—	—
Deep autoencoder [35]	Deep	88.80	—	—
DenseNet121 (baseline)	Deep	94.22 ± 2.10	94.16	93.94
DenseNet169 (baseline)	Deep	94.85 ± 1.12	94.42	94.15
DenseNet201 (baseline)	Deep	94.16 ± 1.79	94.17	94.02
DenseNet264 (baseline)	Deep	93.27 ± 1.64	93.07	93.17
ILQP + ML-DenseNet (proposed approach)	Handcrafted + deep	95.59 ± 1.02	95.61	95.28