Table 3 Texture retention analysis in No RNN and Full Model pipelines.

From: Region of interest-specific loss functions improve T2 quantification with ultrafast T2 mapping MRI sequences in knee, hip and lumbar spine

  

R

GLCM texture metric

Contrast

Dissimilarity

Homogeneity

ASM

Energy

Knee

Full model

2

0.307 ± 0.18**

0.638 ± 0.12***

0.734 ± 0.09***

0.966 ± 0.015***

0.954 ± 0.02***

3

0.153 ± 0.2

0.521 ± 0.15***

0.735 ± 0.09***

0.962 ± 0.015***

0.95 ± 0.02***

4

0.11 ± 0.2

0.387 ± 0.17***

0.61 ± 0.12***

0.973 ± 0.01***

0.95 ± 0.02***

6

0.0667 ± 0.2

0.22 ± 0.19*

0.382 ± 0.17***

0.97 ± 0.015***

0.94 ± 0.025***

8

0.061 ± 0.2

0.111 ± 0.2

0.0615 ± 0.2

0.952 ± 0.02***

0.9 ± 0.04***

10

0.0594 ± 0.2

0.218 ± 0.19*

0.307 ± 0.18**

0.961 ± 0.015***

0.928 ± 0.03***

12

0.0032 ± 0.2

− 0.066 ± 0.2

− 0.178 ± 0.19

0.927 ± 0.03***

0.861 ± 0.055***

No RNN

2

0.455 ± 0.16***

0.599 ± 0.13***

0.32 ± 0.18***

0.898 ± 0.04***

0.904 ± 0.04***

3

0.394 ± 0.17***

0.523 ± 0.15***

0.383 ± 0.17***

0.709 ± 0.11***

0.802 ± 0.07***

4

0.262 ± 0.18**

0.305 ± 0.18**

0.244 ± 0.18**

0.646 ± 0.12***

0.754 ± 0.09***

6

0.103 ± 0.2

0.0574 ± 0.2

0.061 ± 0.2

0.874 ± 0.045***

0.869 ± 0.05***

8

0.0645 ± 0.2

0.0411 ± 0.2

0.0435 ± 0.2

0.922 ± 0.03***

0.911 ± 0.035***

10

0.0474 ± 0.2

0.0382 ± 0.2

0.093 ± 0.2

0.92 ± 0.035***

0.913 ± 0.035***

12

0.0568 ± 0.2

0.0315 ± 0.2

0.0885 ± 0.2

0.818 ± 0.065***

0.862 ± 0.055***

Hip

Full model

2

0.312 ± 0.34*

0.633 ± 0.23***

0.837 ± 0.12***

0.945 ± 0.04***

0.957 ± 0.035***

3

0.369 ± 0.32*

0.671 ± 0.21***

0.816 ± 0.14***

0.976 ± 0.02***

0.98 ± 0.015***

4

0.328 ± 0.33*

0.597 ± 0.25***

0.801 ± 0.15***

0.957 ± 0.035***

0.954 ± 0.04***

6

0.235 ± 0.35

0.475 ± 0.3**

0.645 ± 0.23***

0.939 ± 0.05***

0.941 ± 0.045***

8

0.199 ± 0.36

0.487 ± 0.28**

0.823 ± 0.13***

0.923 ± 0.06***

0.933 ± 0.055***

10

0.127 ± 0.36

0.308 ± 0.34

0.48 ± 0.29**

0.862 ± 0.11***

0.855 ± 0.11***

12

0.198 ± 0.36

0.38 ± 0.32*

0.523 ± 0.28**

0.927 ± 0.06***

0.914 ± 0.07***

No RNN

2

0.285 ± 0.34

0.399 ± 0.32*

0.406 ± 0.31*

0.855 ± 0.11***

0.841 ± 0.12***

3

0.15 ± 0.36

0.241 ± 0.35

0.292 ± 0.34

0.867 ± 0.1***

0.85 ± 0.12***

4

0.113 ± 0.36

0.202 ± 0.36

0.282 ± 0.34

0.836 ± 0.12***

0.813 ± 0.14***

6

0.0394 ± 0.36

0.0504 ± 0.36

0.0785 ± 0.36

0.793 ± 0.15***

0.767 ± 0.16***

8

0.0229 ± 0.37

0.000593 ± 0.37

− 0.0583 ± 0.37

0.682 ± 0.21***

0.653 ± 0.22***

10

− 0.00292 ± 0.36

− 0.0328 ± 0.37

− 0.196 ± 0.36

0.644 ± 0.23***

0.621 ± 0.24***

12

− 0.00208 ± 0.37

− 0.0312 ± 0.36

− 0.0646 ± 0.36

0.712 ± 0.2***

0.687 ± 0.2***

Lumbar spine

Full model

2

0.557 ± 0.7

0.695 ± 0.62

0.744 ± 0.57*

0.892 ± 0.35**

0.923 ± 0.27**

3

0.499 ± 0.73

0.615 ± 0.67

0.644 ± 0.66

0.819 ± 0.48*

0.872 ± 0.39*

4

0.236 ± 0.8

0.421 ± 0.76

0.497 ± 0.73

0.67 ± 0.64

0.775 ± 0.54*

6

0.341 ± 0.78

0.428 ± 0.76

0.262 ± 0.8

0.566 ± 0.7

0.67 ± 0.64*

8

0.0633 ± 0.81

0.152 ± 0.8

0.276 ± 0.79

0.685 ± 0.62

0.728 ± 0.58*

10

− 0.0393 ± 0.81

− 0.0631 ± 0.81

− 0.0699 ± 0.81

0.403 ± 0.76

0.479 ± 0.74*

12

− 0.0697 ± 0.81

− 0.156 ± 0.8

− 0.424 ± 0.76

0.16 ± 0.8

0.198 ± 0.8*

No RNN

2

0.496 ± 0.73

0.731 ± 0.58*

0.883 ± 0.37**

0.967 ± 0.14***

0.975 ± 0.11***

3

0.357 ± 0.78

0.615 ± 0.67

0.807 ± 0.5*

0.909 ± 0.31**

0.934 ± 0.24**

4

0.336 ± 0.78

0.607 ± 0.68

0.771 ± 0.54*

0.874 ± 0.38*

0.91 ± 0.31**

6

0.307 ± 0.78

0.53 ± 0.72

0.604 ± 0.68

0.903 ± 0.32**

0.916 ± 0.29**

8

0.2 ± 0.8

0.4 ± 0.76

0.59 ± 0.68

0.847 ± 0.44*

0.871 ± 0.39*

10

0.0696 ± 0.81

0.184 ± 0.8

0.386 ± 0.76

0.692 ± 0.62

0.726 ± 0.59*

12

0.0157 ± 0.82

0.0858 ± 0.81

0.325 ± 0.78

0.561 ± 0.7

0.591 ± 0.68*

  1. Intraclass correlation coefficients (ICCs) of Gray Level Co-Occurrence Matrix (GLCM)-based metrics. Contrast and dissimilarity are most sensitive to sharper image textures, while homogeneity, ASM, and energy are most sensitive to smoother image textures. Significance in correlations is noted as follows: *P < 0.05, **P < 0.01, ***P < 0.001 (knee: n = 16; hip: n = 15; lumbar spine: n = 5). In the knee and hip, Full Model pipelines outperformed No RNN versions in retention of smooth and sharp textures. In the lumbar spine, the No RNN pipeline outperformed the Full Model version, possibly because the smaller lumbar spine dataset size made training a larger network with a multi-component loss more difficult. In conjunction with standard reconstruction metrics, the Full Model pipeline was selected as the best knee and hip model, whereas the No RNN pipeline was selected as the best lumbar spine model. Top models in all anatomies preserved smoother textures at nearly all tested R, while dissimilarity texture metrics showed sharper textures were significantly correlated with ground truth and preserved in the knee and hip at low to medium R. In the lumbar spine, mean ICCs for sharper textures at many tested R also were high, but small dataset size likely led to wide standard deviations that prevented significant conclusions from being reached. All told, many textures are preserved in T2 maps by all pipelines, particularly in the knee and hip.
Back to article page