TY  - JOUR
T1  - Lung Cancer Detection by Automatic Region Growing with Morphological
Masking and Neural Network Classifier
AU - Manikandan, T. AU - Bharathi, N. 
JO  - Asian Journal of Information Technology
VL  - 15
IS  - 21
SP  - 4189
EP  - 4194
PY  - 2016
DA  - 2001/08/19
SN  - 1682-3915
DO  - ajit.2016.4189.4194
UR  - https://makhillpublications.co/view-article.php?doi=ajit.2016.4189.4194
KW  - Lung cancer
KW  -lung nodule
KW  -region growing
KW  -morphological masking
KW  -neural network
AB  - An effective lung cancer detection scheme is developed in the present research for segmenting the suspected nodules from the consecutive slices of Computer Tomography (CT) images using Automatic Region Growing (ARG) with morphological masking. The seed values are chosen automatically for the region growing technique. After the initial segmentation of the suspected lung nodules by the region growing technique, features like eccentricity, area and convex area were computed to eliminate unwanted line like structures and small tiny non-cancerous nodules, (>3 mm in size). Further, the centroid shift for the remaining suspected nodules in consecutive slices were calculated to reduce the false positives (vessels), as the centroid shift do not vary much for cancerous nodules and calcifications (calcium deposition on the lungs). Finally, the texture features like homogeneity, auto-correlation and contrast were calculated to remove the calcifications. The extracted features like centroid shift, homogeneity, auto-correlation and contrast were utilized to train and test the neural network classifier. The study was carried out on a total of 106 patients CT scan images (56 cancerous patients and 50 non-cancerous subjects) retrospectively collected from the Bharat Scans, Chennai, which had 60 malignant (cancerous) nodules and 238 benign (non-cancerous) nodules. Out of these nodules, 56 true malignant nodules and 204 true benign nodules were classified correctly by the neural network classifier and remaining was misclassified. Present work produced good sensitivity and specificity and accuracy of 93, 86 and 87%, respectively. The False Positive (FP) per patient is reported as 0.32.
ER  -