Page 31 - TD-2-1

Tumor Discovery An approach for classification of lung nodules

6.1. Lung nodule segmentation 7. Feature extraction

In region growing, the procedure called labeling is done to The diameter, area, irregularity index, and perimeters are
put the negative number representing the label of the region some of the geometrical features that will be evaluated from
to which the pixel has a place. The labeling process keeps the separated lung nodules. The number of pixels in the
tracks of the record of pixels which are yet to be replaced picture array with the value 1 will be used to determine the size
by the labeling. With this reference list, the operation of of the segmented tumor image. The area is estimated using
insertion and removal will be carried out. Removal is done the established technique using bit quads, or 2-by-2pixel
by eliminating the pixels from the front list, and insertion patterns. The quantity of boundary pixels provides an
is by inserting the pixels at the end of the list. estimate of the tumor image’s perimeter. In terms of its

A B C morphology, the tumor has a circular form. The circulatory
index will be calculated to identify the irregularity in the
circular shape using the expression I = 4πA/p2, where A is
the area of the tumor and P is the perimeter of the tumor in
pixels. Table 2 shows the geometrical features for Figure 5.
The important features used in the lung cancer classification
are texture or contrast features.

st
nd
Figure 5. (A) Original image, (B) separated lung fields, and (C) separated The 1 order statistic and 2 order statistic are the
cancerous portion categories under which the contrast features are classified.
The shorter processing time and lower minimum cost are
Table 2. Calculated SMF of the filtered images for all types the advantages in feature extraction using wavelet due to
of filters the solid depiction of the wavelet transform.
Filter type Output SMF 8. Classification of lung nodule
Average filter 7.3133
Weighted average filter 4.2478 Classification is the process of determining whether the
nodules belong to a specified class or not. In supervised
Gaussian filter 6.4443 classification, samples and anticipated classes are known
Selective median filter 8.5816 before a classifier is trained on a set of data. It is evident
Wavelet filter 8.8732 that supervised classification-based methods consistently
[43]
Wiener filter 13.3969 outperform other types of traditional classification methods .
SMF: Selective median filter Classification of nodule in the lung is accomplished using

Table 3. Performance metrics of solid feature for various inertia weights
Classifi ers Accuracy y Sensiti vity Specificit y PPV NPV F‑measu re G‑mean MC
DS 88 92 84 85.18 91.30 87.81 87.9 0.76
RF 88 92 84 85.18 91.30 87.81 87.9 0.76
Ada-DS 86 92 80 82.14 90.9 85.58 85.79 0.72
Ada-RF 90 100 80 83.33 100 88.89 89.44 0.81
BPNN 98 96 100 100 96.15 97.95 97.97 0.96
PPV: Positive predictive value, NPV: Negative predictive value, DS: Decision Stump, RF: Random Forest, BPNN: Back Propagation Neural Network
Table 4. Performance metrics of part‑solid feature for various inertia weights

Classifie rs Accura cy Sensitiv ity Specifi city PPV NPV F‑measur e G‑mean C
DS 77.6 60.19 95.14 92.53 70.5 73.54 76.17 0.59
RF 86.2 85.4 86.4 86.27 85.57 85.49 85.49 0.72
Ada-DS 87.9 83.5 92.2 91.48 84.82 86.81 87.74 0.61
Ada-RF 86.4 85.4 87.4 91.48 84.82 87.89 86.39 0.76
BPNN 93.68 91.26 96.1 95.91 96.11 93.43 93.46 0.87
PPV: Positive predictive value, NPV: Negative predictive value, DS: Decision Stumpe, RF: Random Forest, BPNN: Back Propagation Neural Network

Volume 2 Issue 1 (2023) 6 https://doi.org/10.36922/td.317