Pulmonary Nodule Computed Tomography Image Classification Method Based on Dual-Path Cross-Fusion Network

Table 1. Parameter setting
View table
Table 1. Parameter setting
Training parameter Parameter value
Batch size 64
Train epochs 50
Optimizer AdamW
Learning rate 10^-4
Learning rate schedule Cosine annealing
Optimizer momentum β₁，β₂=0.9，0.9999
Weight decay 0.01

Table 2. Comparison of classification metrics of different models

View table

Table 2. Comparison of classification metrics of different models

Model	Accuracy	Recall	Precision	Specificity	AUC
ResNet50	86.21	83.99	87.04	88.25	91.28
DenseNet121	84.91	81.00	86.69	88.40	90.75
ConvNeXt-Base	89.90	89.70	89.48	90.00	94.60
ViT-Base	90.44	89.66	90.36	91.18	94.74
Swin Transformer-Base	90.34	89.50	90.32	91.10	94.59
Proposed method	94.16	93.93	93.03	92.54	97.02

Table 3. Comparison of parameter number, FLOPs and inference time of different models
View table
Table 3. Comparison of parameter number, FLOPs and inference time of different models
Model Parameter /M FLOPs /G Inference time /ms
ResNet50 23.5 4.1 13.7
DenseNet121 7.0 2.9 19.3
ConvNeXt-Base 87.6 15.4 15.3
ViT-Base 85.8 16.9 13.8
Swin Transformer-Base 86.7 15.2 19.9
Proposed method 28.2 4.0 10.3

Table 4. Comparison with different research methods

View table

Table 4. Comparison with different research methods

Type	Method	Accuracy	Recall	Precision	Specificity	AUC
CNN	Ref.［23］	88.46	88.66	87.38	—	95.62
	Ref.［24］	90.77	85.37	—	95.04	—
	Ref.［25］	92.81	92.36	92.59	—	96.17
	Ref.［26］	91.07	90.93	91.18	91.22	95.84
	Ref.［27］	91.25	89.10	91.59	93.39	91.25
Transformer	Ref.［12］	93.33	—	—	—	96.04
CNN+Transformer	Ref.［10］	92.92	93.84	91.62	—	96.28
CNN+Transformer	Proposed method	94.16	93.93	93.03	92.54	97.02

Table 5. Results of the ablation experiment

View table

Table 5. Results of the ablation experiment

Model	Local path	Global path	Feature fusion block	Hybrid loss function	Accuracy	Recall	Precision	Specificity	AUC
M1	√				85.42	81.05	86.54	86.38	90.82
M2		√			85.26	83.97	85.42	86.43	90.56
M3	√			√	86.45	85.66	86.36	87.18	91.06
M4		√		√	87.39	86.86	86.44	86.95	91.61
M5	√	√			90.89	89.67	90.93	91.62	95.31
M6	√	√	√		92.04	92.61	90.56	91.65	96.23
M7	√	√	√	√	94.16	93.93	93.03	92.54	97.02

Tools

Get Citation

Copy Citation Text

Ping Yang, Xin Zhang, Fan Wen, Ji Tian, Ning He. Pulmonary Nodule Computed Tomography Image Classification Method Based on Dual-Path Cross-Fusion Network[J]. Laser & Optoelectronics Progress, 2024, 61(8): 0837002

Download Citation

EndNote(RIS)BibTex Plain Text

Set citation alerts for article

Save article for my favorites