Unsupervised detection of density changes through principal component analysis for lung lesion classification

Abstract

Lung cancer remains one of the most common cancers worldwide. Temporal evaluation is a useful tool for analyzing the malignant behavior of a lesion during treatment or that of indeterminate lesions which may be benign. Thereby, this work proposes a methodology for analysis, quantification and visualization of unsupervised changes in lung lesions, through principal component analysis. From change regions, we extracted texture features for lesion classification as benign or malignant. To reach this purpose, two databases with distinct behavior were used, one of which concerning malign under treatment and another indeterminate, but likely benign, lesions. The results have shown that the lesion’s density changes in a public database of malignant lesions under treatment were greater than the private database of benign lung nodules. From the texture analysis of the regions where the density changes occurred, we were able to discriminate lung lesions with an accuracy of 98.41 %, showing that these changes could point out the nature of the lesion. Other contribution was visualization of changes occurring in the lesions over time. Besides, we quantified these changes and analyzed the entire set through volumetry, the most commonly used technique to evaluate progression of lung lesions.

Appendix: Texture Analysis

Texture features were used in this work for the diagnostic classification task of lung lesions.

1.1 Run-Length Features

These capture the gray level primitives’ proprieties. M is the run-length matrix, a is the gray level value and r is the length of primitive.

1.1.1 Short-Run Emphasis - (SRE)

$$ SRE=\frac{1}{K}\sum\limits_{a=0}^{G-1}\sum\limits_{r=1}^{N_{r}}\frac{M(a,r)}{r^{2}} $$

(2)

1.1.2 Long-Run Emphasis - (LRE)

$$ LRE=\frac{1}{K}\sum\limits_{a=0}^{G-1}\sum\limits_{r=1}^{N_{r}}M(a,r)r^{2} $$

(3)

1.1.3 Gray-Level Nonuniformity - (GLNU)

$$ GLNU=\frac{1}{K}\sum\limits_{r=1}^{N_{r}}\left[\sum\limits_{a=0}^{G-1}M(a,r) \right]^{2} $$

(4)

1.1.4 Run-Length Nonuniformity - (RLN)

$$ RLN=\frac{1}{K}\sum\limits_{a=0}^{G-1}\left[\sum\limits_{r=1}^{N_{r}}M(a,r) \right]^{2} $$

(5)

1.1.5 Low Gray Level Run Emphasis - (LGRE)

$$ LGRE=\frac{1}{K}\sum\limits_{r=1}^{N_{r}}\sum\limits_{a=0}^{G-1}\frac{M(a,r)}{\left( a + 1 \right)^{2}} $$

(6)

1.1.6 High Gray Level Run Emphasis - (HGLRN)

$$ HGLRN=\frac{1}{K}\sum\limits_{r=1}^{N_{r}}\sum\limits_{a=0}^{G-1}M(a,r)\left( a + 1 \right)^{2} $$

(7)

1.1.7 Short Run, Low Gray Level Emphasis - (SRLGLE)

$$ SRLGLE=\frac{1}{K}\sum\limits_{a=0}^{G-1}\sum\limits_{r=1}^{N_{r}}\frac{M(a,r)}{r^{2}\left( a + 1 \right)^{2}} $$

(8)

1.1.8 Short Run, High Gray Level Emphasis - (SRHGLE)

$$ SRHGLE=\frac{1}{K}\sum\limits_{a=0}^{G-1}\sum\limits_{r=1}^{N_{r}}\frac{M(a,r)\left( a + 1 \right)^{2}}{r^{2}} $$

(9)

1.1.9 Long Run, Low Gray Level Emphasis - (LRLGLE)

$$ LRLGLE=\frac{1}{K}\sum\limits_{a=0}^{G-1}\sum\limits_{r=1}^{N_{r}}\frac{M(a,r)r^{2}}{\left( a + 1 \right)^{2}} $$

(10)

1.1.10 Long Run, High Gray Level Emphasis - (LRHGLE)

$$ LRHGLE=\frac{1}{K}\sum\limits_{a=0}^{G-1}\sum\limits_{r=1}^{N_{r}}M(a,r)r^{2}\left( a + 1 \right)^{2} $$

(11)