Page 48 - IJOCTA-15-4

Kazemi et al. / IJOCTA, Vol.15, No.4, pp.578-593 (2025)
Table 1. Comparison of key characteristics between tension field and 2-tension field.

Aspect Tension Field 2-Tension Field
Linearity/Nonlinearity Linear Nonlinear
Edge detection Highlights edges but may Enhances edges adaptively, reducing
amplify noise noise amplification
Noise sensitivity High sensitivity to noise Less sensitive to noise due to adap-
due to second derivatives tive modulation
Smoothing behavior Isotropic smoothing Anisotropic smoothing: preserves
edges while smoothing within homo-
geneous regions
Complexity Simple and computation- More complex due to additional
ally efficient terms involving gradient magni-
tudes
Use cases Suitable for simple edge Ideal for advanced image processing
detection or preprocessing tasks like denoising, edge-preserving
steps smoothing

Table 2. Overview of the experimental setup.

Dataset Noise Variance Samples Repetitions Experiments
0.01 40 10 400
LIDC-IDRI 0.05 40 10 400
0.1 40 10 400

spatial diversity in the magnitude of the gradient Measure (SSIM) was used to assess the quality of
itself, thus seriously contributing to the stabiliza- the image restoration after denoising. 26
tion and enhancement of high-resolution struc- Table 3 provides an overview of the experi-
tures. In numerical implementation, stability and mental setup.
computational efficiency are achieved through the
Table 3 compares two denoising methods (the
use of finite differences and iterative methods. In 2-tension field operator and the tension field oper-
summary, the 2-tension field represents a signifi- ator) on the LIDC-IDRI dataset at three different
cant leap forward compared to traditional meth- Gaussian noise levels (variance = 0.01, 0.05, 0.1).
ods, as it offers a mathematically sound and The methods are compared based on SSIM scores
flexible approach to surmount the smoothness-
to determine which denoising method preserves
structural preservation trade-off in a variety of
the structural similarity after reducing noise lev-
image processing tasks.
els. For the lower noise levels (variance = 0.01),
the 2-tension field operator had a slightly bet-
ter performance than the tension field operator
5.1. 2-Tension versus tension field (SSIM values = 0.93 vs. 0.92). As noise variance
operator: Gaussian noise reduction in increases (variance = 0.05 and 0.1), the 2-tension
medical imaging
field operator continued to show better perfor-
This part is dedicated to comparing two noise re- mance than the tension field operator (SSIM val-
duction methods–the 2-tension field operator and ues = 0.87 vs. 0.85 and 0.81 vs. 0.76). These
the tension field operator–in terms of how much results show that the 2-tension field operator has
Gaussian noise is diminished. This comparison an increased robustness factor, which preserves
is carried out on the lung CT images for cancer structural integrity under higher noise conditions
detection in the LIDC-IDRI Dataset. 8 than the tension field operator.
Stratified random sampling was used to ob-
tain a representative and diverse sample. 25 Forty
Table 3. SSIM scores for the LIDC-IDRI Dataset
images were randomly sampled from the LIDC-
IDRI dataset, with cases of mixed nodule com-
Noise Variance 2-Tension Tension SSIM
plexity. Each image was corrupted with three
SSIM
amounts of Gaussian noise with variances of 0.01,
0.01 0.93 0.92
0.05, and 0.1. To maintain consistency for evalua-
0.05 0.87 0.85
tion, each experiment was repeated independently
0.1 0.81 0.76
a total of 10 times. Structural Similarity Index
590