Lightweight Directional Transformer for X-Ray-Aided Pneumonia Diagnosis

Tao Zhou; Xinyu Ye; Fengzhen Liu; Huiling Lu

doi:10.3788/AOS230447

Acta Optica Sinica, Volume. 43, Issue 13, 1310001(2023)

Lightweight Directional Transformer for X-Ray-Aided Pneumonia Diagnosis

Tao Zhou^1,2, Xinyu Ye^1,2、*, Fengzhen Liu^1,2, and Huiling Lu³

Author Affiliations

¹School of Computer Science and Engineering, North Minzu University, Yinchuan 750021, Ningxia, China

²The Key Laboratory of Images and Graphics Intelligent Processing of State Ethnic Affairs Commission, North Minzu University, Yinchuan 750021, Ningxia, China

³School of Medical information and Engineering, Ningxia Medical University, Yinchuan 750004, Ningxia, China

show less

Abstract Get PDF(in Chinese)

Figures & Tables(11)

Fig. 1. Overall frame structure of LDTransformer

Download full size

Fig. 2. Structure of directional Transformer. (a) Directional Transformer layer; (b) directional convolution; (c) calculation process

Download full size

Fig. 3. Structure of lightweight convolution. (a) General convolution; (b) lightweight convolution

Download full size

Fig. 4. Thermograms of ablation model on pneumonia X-ray

Download full size

Fig. 5. ROC curves of each model

Download full size

Fig. 6. P-R curves of each model

Download full size

Table 1. Specific results of ablation experiments

View table

Table 1. Specific results of ablation experiments

Model	Parameter amount	Calculation amount	Total time /s	A /%	AUC /%	R /%	F₁ /%	P /%
Base	6.96×10⁶	2.87×10⁹	16670	93.44±2.1	93.60±2.0	92.69±1.6	94.33±1.7	96.04±1.8
1	5.09×10⁶	2.18×10⁹	17285	94.30±3.3	94.26±3.5	94.49±3.4	95.14±3.3	95.78±3.2
2	4.68×10⁶	1.81×10⁹	16381	95.83±2.3	95.63±2.2	96.74±2.2	96.47±2.3	96.19±2.4
3	3.75×10⁶	1.27×10⁹	15564	95.49±2.7	95.44±2.8	95.73±2.8	96.16±2.7	96.59±2.6
4	1.26×10⁶	4.95×10⁸	13479	97.22±2.0	97.17±2.3	97.41±2.1	97.63±1.9	97.85±2.0
5	2.53×10⁵	3.98×10⁷	12047	97.54±1.7	97.53±1.9	97.64±1.7	97.91±1.6	98.19±1.6

Table 2. Comparison of evaluation indices with different loss functions

View table

Table 2. Comparison of evaluation indices with different loss functions

Model	A /%	AUC /%	R /%
Cross entropy loss	97.54±1.7	97.53±1.9	97.64±1.7
Hinge loss^［15］	97.88±0.9	97.86±1.1	97.98±1.0
Symmetric cross entropy loss^［16］	98.01±1.7	97.99±1.9	98.09±1.8
Focal loss^［16］	98.41±1.9	98.43±2.0	98.31±1.9
Class balance focal loss^［19］	98.48±1.5	98.51±1.7	98.31±1.8
Balanced focal loss	98.87±0.8	98.85±1.0	98.99±0.9

Table 3. Comparison results of each model in pneumonia X-ray dataset

View table

Table 3. Comparison results of each model in pneumonia X-ray dataset

Model	Parameter amount	Calculation amount	Total time /s	A /%	AUC /%	R /%	F₁ /%	P /%
EfficientNetb0^［3］	4.48×10⁴	1.35×10⁷	15155	91.52±2.5	91.68±2.8	90.78±3.1	92.65±2.6	94.61±2.4
MobileNetV3^［13］	1.66×10⁶	6.24×10⁷	14757	92.05±1.8	92.03±2.3	92.13±2.2	93.17±1.9	94.25±1.8
GhostNet^［5］	3.90×10⁶	1.48×10⁸	15398	94.17±2.0	94.05±2.2	94.71±2.4	95.03±2.1	95.36±2.0
DeiT-S^［8］	2.19×10⁶	4.24×10⁹	21986	92.78±3.3	92.63±3.6	93.48±3.1	93.85±3.2	94.22±3.3
T2T-ViT-19^［9］	2.14×10⁶	4.33×10⁹	22179	94.62±2.6	94.37±2.9	95.84±2.4	95.46±2.5	95.09±2.5
XCiT-S24-16T^［20］	4.76×10⁷	8.95×10⁹	25142	96.02±1.9	95.94±2.1	96.40±1.8	96.62±1.8	96.84±1.9
ConvViT^［10］	4.22×10⁵	9.08×10⁷	15817	93.17±3.9	92.99±3.8	94.04±4.1	94.20±3.8	94.36±4.0
Mobile-Former^［13］	4.58×10⁶	9.26×10⁷	16124	93.37±2.8	93.28±2.7	93.81±2.8	94.43±2.7	94.88±2.9
Mobile-ViT^［14］	5.63×10⁶	1.75×10⁹	18978	95.19±2.3	95.19±2.3	96.00±2.4	95.27±2.3	94.55±2.2
EdgeNeXt-S^［21］	5.58×10⁶	9.57×10⁸	16535	96.02±2.5	95.92±2.4	96.51±2.3	96.62±2.4	96.73±2.6
NextVit-S^［22］	3.18×10⁷	5.79×10⁹	20136	96.75±2.1	96.63±2.0	97.30±2.2	97.25±2.1	97.19±2.0
LDTransformer	2.53×10⁵	3.98×10⁷	11783	98.87±0.8	98.85±1.0	98.99±0.9	99.04±0.9	99.10±0.8

Table 4. Results of the first public dataset

View table

Table 4. Results of the first public dataset

Model	TPR /%	TNR /%	A /%	AUC /%
DenseNet121^［12］	91.00	87.00	88.13	90.00
EfficientNetb0^［3］	83.00	92.00	94.64	95.00
Covid-caps^［23］	90.00	95.00	95.00	97.00
ViT-B32^［23］	96.00	96.00	96.00	99.10
LDTransformer	98.55	98.97	98.71	99.53

Table 5. Results of the second public dataset
View table
Table 5. Results of the second public dataset
Model A /% AUC /%
CVDNet^［24］ 97.20 -
AF-CAP^［24］ 99.16 98.80
LDTransformer 99.63 99.17

Tools

Get Citation

Copy Citation Text

Tao Zhou, Xinyu Ye, Fengzhen Liu, Huiling Lu. Lightweight Directional Transformer for X-Ray-Aided Pneumonia Diagnosis[J]. Acta Optica Sinica, 2023, 43(13): 1310001

Download Citation

EndNote(RIS)BibTex Plain Text

Set citation alerts for article

Save article for my favorites