Aims and objectives
Chest X-Ray (CXR) is one of the most commonly performed radiological exams.
Various computer-assisted diagnostic (CAD) algorithms have already been proposed for automated CXR pathology classification.
However,
the use of CXR CAD tools is not yet widespread in a clinical setting.
An ideal algorithm should differentiate between a variety of CXR pathologies that manifest similarly in appearance and be robust with respect to different CXR images.
Using a large and inhomogeneous dataset from 5 medical centers we aim to build an end-to-end model for multiclass...
Methods and materials
A database of CXR images (n = 116132) from 5 medical centers from 3 continents was used (listed in terms of quantity): Chest-xray8 [3],
Vilnius University hospital Santaros clinics,
openi [4],
tuberculosis_by [5],
tuberculosis_India [6],
tuberculosis_Montgomery [7],
tuberculosis_Shenzhen [8] .
To our knowledge,
this is the biggest CXR database reported in the context of a CXR CAD to date.
We used labels that were either provided with the images (Chest-xray8,
openi,
tuberculosis_by,
tuberculosis_India,
tuberculosis_Montgomery,
tuberculosis_Shenzhen) or text mined from the associated radiologist descriptions (Vilnius University...
Results
The obtained under receiver operating characteristic curve (AUC) testing scores for different pathologies (each versus the rest) are listed in table 1:
Table 1:
Label
Freq
AUC
No Finding
0.0815
0.8582
Cardiomegaly
0.0027
0.7608
Effusion
0.7385
0.8198
Atelectasis
0.2683
0.6241
Pneumothorax
0.0751
0.6925
Mass/Nodule
0.0039
0.7266
Infiltration
0.1330
0.6169
Pleural Thickening
0.0045
0.6446
Emphysema
0.029
0.7912
Fibrosis
0.0661
0.7099
Edema
0.0231
0.7695
Tuberculosis
0.0574
0.4790
Average AUC with(out) TB
0.7077 (0.7285)
Highest AUC scores were obtained for labels no finding,
pleural effusion,
emphysema,
edema and...
Conclusion
We have demonstrated robust CXR classification results on a particularly heterogeneous set of radiographs from different medical centers worldwide.
We have also created an API to clearly visualize the neural network output and the localization of the detected labels.
Using data from different medical centers contributes greatly to the data heterogeneity and pose significant challenges to creating a neural network which could provide high AUC scores for pathology detection. This however is necessary while working towards a robust algorithm that could be implemented into a...
Personal information
N.
Ramanauskas -Department of Radiology,
Nuclear Medicine and Medical Physics,Institute of Biomedical Sciences,
Faculty of Medicine,
Vilnius University,Vilnius University Hospital Santaros Klinikos; Oxipit,
UAB
J.
Dementavičienė -Department of Radiology,
Nuclear Medicine and Medical Physics,Institute of Biomedical Sciences,
Faculty of Medicine,
Vilnius University,
Vilnius University Hospital Santaros Klinikos
J.
Bialopetravičius -Institute of Theoretical Physics and Astronomy,
Vilnius University,
Vilnius,
Lithuania; Oxipit,
UAB
D.
Barušauskas - Oxipit,
UAB
J.
Armaitis - Oxipit,
UAB
J.
Stankevičienė -Department of Radiology,
Nuclear Medicine and Medical Physics,Institute of Biomedical Sciences,
Faculty...
References
1.
Esteva A,
Kuprel B,
Novoa RA,
et al.
Dermatologist-level classification of skin cancer with deep neural networks.
Nature 2017;542:115-118
2.
Gulshan V,
Peng L,
Coram M,
et al.
Development and Validation of a Deep Learning Algorithm for Detection of Diabetic Retinopathy in Retinal Fundus Photographs.
JAMA 2016;316(22):2402-2410
3.
Wang X,
Peng Y,
Lu L,
et al.
ChestX-ray8: Hospital-scale Chest X-ray Database and Benchmarks on Weakly-Supervised Classification and Localization of Common Thorax Diseases.
IEEE CVPR 2017;3462-3471
4.
https://nihcc.app.box.com/v/ChestXray-NIHCC/folder/36938765345
5.
D.
Demner-Fushman D ,
M.
D....