Page 33 - TD-2-1

Tumor Discovery An approach for classification of lung nodules

precision of a positive class is defined by sensitivity. In 10. Conclusions
addition, specificity determines how well a negative class is
classified. The value nearer to 100% represents the perfect In this work, we proposed an automated CAD system for
classification accuracy. classification of lung nodules using various classifiers from
CT images. The classification of nodule and non-nodule
*
Gmean sensitivity specificity (VII) patterns in CT is one of the most significant processes
during the detection of lung nodule. The developed
CAD systems consist of segmentation, feature extraction
8.10. Mathew’s correlation
and classification. For segmentation, we used filters for
In Mathew’s correlation (MC), the actual and predicted effective extraction infected region. Later, we extracted
condition takes the value between 0 and 1. The value of 1 features through features and fed into classifiers such as
corresponds to perfect correlation, whereas the value of 0.5 DS, RF, and BPNN. The experimentation was conducted
corresponds to random prediction. on LIDC-IDRI dataset (Tables 3-5), and the results with
TP TN* ( FP FN* ) (VIII) BPNN outperformed those with DS and RF classifiers. The
performance was measured using sensitivity, specificity,
FP TN
TP FP TP FN TN FN / 12 PPV, NPV, F-measure, and G-Mean.
TP TN* FP FN* Acknowledgments

AP AN PP PN* * * / 12 None.

9. Results and discussion Funding
None.
The images used for examining the proposed methodology
were taken from the LIDC-IDRI, SPIE-AAPM Lung CT Conflict of interest
challenge, and hospitals. Nodule size between 3 mm and
30 mm were considered in this work. Most specifically, The authors declare no conflict of interest.
solid, part-solid, and non-solid nodules were chosen. Author contributions
In the LIDC-IDRI database, 71 exams are chosen. Out
of 71 exams, 246 nodule case and 240 non-nodule cases Conceptualization: Naveen HM
were selected. In the SPIE-AAPM database, out of 70 only Investigation: Naveen HM
35 exams were used. Among them, 28 nodule cases and Methodology: Naveen HM
34 non-nodule cases are selected. About 36 CT images Formal analysis: Naveena C, Manjunath Aradhya VN
were acquired from hospitals. A total of 584 images were Writing – original draft: Naveen HM
considered in this work, of which 292 belong to nodule and Writing – review & editing: Naveen HM
the rest belong to non-nodule cases. The input datasets are Ethics approval and consent to participate
grouped into training set and testing set with 292 datasets
each. All these databases were aimed to promote the Not applicable.
development of the proposed CAD system.
Consent for publication
The performance measure of each classifier for different
inertia weights can be measure using accuracy, sensitivity, Not applicable.
and specificity. Using confusion matrix, these measures Availability of data
can be calculated. Accuracy of each classifier can be
obtained correctly by determining the ratio of the correctly Not applicable.
classified and total number of samples. Sensitivity can be References
measured from the misclassified rate of nodule case to
the total number of nodule case used. Specificity can be 1. Akbari R, Ziarati K, 2011, A rank based particle swarm
measured from the misclassified rate of non-nodule case optimization algorithm with dynamic adaptation. J Comput
to the total number of non-nodule case used. Tables 1-3 Appl Math., 235(8): 2694–2714.
describe briefly the performance of classifiers based https://doi.org/10.1016/j.cam.2010.11.021
on confusion matrix. For solid features, the accuracy, 2. Aoyama M, Li Q, Katsuragawa S, et al., 2003, Computerized
sensitivity, specificity, PPV, NPV, F-measure, G-mean, and scheme for determination of the likelihood measure of
MC of various classifiers are noted. malignancy for pulmonary nodules on low-dose CT images.

Volume 2 Issue 1 (2023) 8 https://doi.org/10.36922/td.317