Keywords:
Infection, Computer Applications-General, Digital radiography, Conventional radiography, Respiratory system, Artificial Intelligence
Authors:
T. Fung1, J. W. Luo2, T. C. Lee3, B. Gallix2, J. J. R. Chong2; 1Montreal, Québec/CA, 2Montreal, QC/CA, 3Montreal/CA
DOI:
10.26044/ecr2019/C-3416
Conclusion
Comparison of Network Performance:
Our trained neural network yielded an AUC of 0.82-0.94 which is comparable to previously published detection statistics of other experimental neural networks.
For instance,
Hwang et al.
(2016) had AUCs ranging from 0.88 to 0.96 [2].
Our study shows that DCNN transfer learning approach is able to identify TB on chest radiographs.
Subjective Evaluations Patterns:
However,
unlike previous studies,
we wish to draw certain additional findings,
particularly from the subjective network evaluation.
From our analyses with the generated test Class Activation Maps,
localization performance seemed to be strongest over infiltrate and consolidation manifestations of disease both with the strongest class activations gradients and the best localization performance with strong conformance of the activated boundaries to underlying findings.
For cavitary manifestations of disease,
even in cases where a correct classification was made,
the strength of gradient activation would be weaker.
From a clinical interpretive standpoint,
this is somewhat surprising given how specifically pulmonary cavitation can imply the presence of TB given the appropriate presentation.
From a CNN perspective however,
it is somewhat understandable that the lucent portion of a cavitation resembles the brightness and texture of normal lung parenchyma which may weaken activation over such findings.
With respect to miliary TB,
we felt that the number of examples available in the training set in conjunction with the limited spatial resolution of the downsampled images likely prevented successful training of this finding.
Conclusion:
Tuberculosis is an issue that implicates health care systems globally and involves mandatory reporting and isolation of active cases.
The delay of detection and treatment of disease results in worse outcomes and can further disease spread within small communities.
Given the importance of timely detection and isolation of active TB cases,
AI can be a useful tool to help triage the high volume of images accurately.
Ultimately,
isolation and expert radiographic review of suspected cases can be expedited,
with the possibility of doing so at a much more scalable and responsive manner than current clinical resources permit.