Background
Methods
Clinical cases
Training group | Testing group | |||
---|---|---|---|---|
Benign lesions#
| 1.3(0.5-3.0)cm | 1.8(0.5-9.0) | ||
Malignant lesions#
| 2.8(1.5-5.0)cm | 2.6(0.5-5.5)cm | ||
Number | Percentage | Number | Percentage | |
BI-RADS assessments
| ||||
category 2 | 30 | 12.8 | 17 | 18.3 |
category 3 | 41 | 17.6 | 41 | 44.1 |
category 4 | 98 | 41.8 | 27 | 29.0 |
category 5 | 65 | 27.8 | 8 | 8.6 |
Malignant lesions
|
149
|
63.68
|
75
|
80.6
|
Invasive ductal carcinoma | 120 | 51.3 | 62 | 66.7 |
Intraductal carcinoma | 17 | 7.26 | 9 | 9.7 |
Ductal carcinoma in situ | 4 | 1.7 | 1 | 1.1 |
Mucinous carcinoma | 3 | 1.28 | 2 | 2.1 |
Medullary carcinoma | 1 | 0.43 | 0 | 0 |
Others | 4 | 1.71 | 1 | 1.1 |
Benign lesions
|
85
|
36.32
|
18
|
19.4
|
Fibroadenoma | 26 | 11.11 | 6 | 6.4 |
Fibrocystic changes | 24 | 10.26 | 3 | 3.2 |
Fibroadenosis | 3 | 1.28 | 3 | 3.2 |
Intraductal papilloma | 4 | 1.71 | 3 | 3.2 |
Hyperplasia | 3 | 1.28 | 1 | 1.1 |
Phyllodestumor | 2 | 0.85’ | 1 | 1.1 |
Adenomyosisepithelioma | 1 | 0.43 | 0 | 0 |
Inflammation | 1 | 0.43 | 1 | 1.1 |
Follow-up | 21 | 8.97 | 0 | 0 |
Image acquisition
Lesion image segmentation
Pictorial characterization of the segmented lesion from MR images
Classification performance of individual features
Classification performance of multi-sided features
Results
Diagnostic performance of each feature individually
Feature name | Parameter distribution*
|
P-value#
| Specificity | Sensitivity | Accuracy | AUC | |
---|---|---|---|---|---|---|---|
Benign | Malignant | ||||||
Elongation
|
0.84 ± 0.13
|
0.87 ± 0.11
|
0.39
|
0.11
|
0.89
|
74.19
|
0.58
|
ADC
|
1.04 ± 0.21
|
1.50 ± 0.43
|
0.00
|
0.67
|
0.92
|
87.10
|
0.85
|
SER
| 1.20 ± 0.22 | 1.00 ± 0.50 | 0.01 | 0.33 | 0.97 | 84.95 | 0.71 |
Correlation
| 0.65 ± 0.16 | 0.60 ± 0.17 | 0.23 | 0.06 | 0.96 | 78.49 | 0.60 |
Inertia
| 1995.34 ± 1177.11 | 2773.68 1891.29 | 0.03 | 0.17 | 0.93 | 78.49 | 0.64 |
Entropy
|
8.48 ± 1.30
|
7.90 ± 1.34
|
0.09
|
0.11
|
0.99
|
81.72
|
0.64
|
Inverse Difference
| 0.10 ± 0.05 | 0.09 ± 0.03 | 0.22 | 0.11 | 0.97 | 80.65 | 0.65 |
Sum average
|
310.21 ± 37.83
|
285.58 ± 34.20
|
0.01
|
0.11
|
0.99
|
81.72
|
0.70
|
Sum variance
|
9235.63 ± 2999.89
|
10078.03 ± 3168.82
|
0.29
|
0.00
|
1.00
|
80.65
|
0.57
|
Sum entropy
| 6.76 ± 0.87 | 6.29 ± 0.90 | 0.04 | 0.11 | 0.99 | 81.72 | 0.66 |
Difference average
| 32.44 ± 10.15 | 38.73 ± 14.30 | 0.03 | 0.28 | 0.93 | 80.65 | 0.66 |
Difference variance
| 820.01 ± 486.86 | 1042.80 636.38 | 0.10 | 0.06 | 0.99 | 80.65 | 0.62 |
Difference entropy
| 5.40 ± 0.44 | 5.16 ± 0.47 | 0.04 | 0.11 | 0.99 | 81.72 | 0.66 |
Information Correlation 1
| −0.58 ± 0.12 | −0.61 ± 0.14 | 0.43 | 0.11 | 0.95 | 78.49 | 0.58 |
Diagnostic performance of multi-sided features in combination
Scenario 1: Entire features outperform individual or combinations of features during diagnostic classification
Classifier | Feature subset | Specificity | Sensitivity | Accuracy | AUC |
---|---|---|---|---|---|
SVM
| Morphology | 0.278 | 0.817 | 67.74 | 0.526 |
Morphology + Texture | 0.444 | 0.851 | 69.89 | 0.602 | |
ADC + SER | 0.722 | 0.926 | 81.72 | 0.781 | |
Morphology + Kinetic | 0.5 | 0.875 | 77.42 | 0.67 | |
Morphology + ADC
|
0.611
|
0.903
|
81.72
|
0.739
| |
Morphology + Texture + Kinetic | 0.556 | 0.882 | 75.27 | 0.678 | |
Entire
*%
|
0.722
30%
|
0.924
4.8%
|
79.57
5.7%
|
0.768
13.3%
| |
KNN
| Morphology | 0.5 | 0.85 | 64.52 | 0.569 |
Morphology + Texture | 0.444 | 0.844 | 66.67 | 0.619 | |
ADC + SER | 0.722 | 0.917 | 73.12 | 0.784 | |
Morphology + Kinetic | 0.556 | 0.867 | 66.67 | 0.66 | |
Morphology + ADC
|
0.611
|
0.892
|
74.19
|
0.794
| |
Morphology + Texture + Kinetic | 0.611 | 0.887 | 70.97 | 0.666 | |
Entire
*%
|
0.611
0%
|
0.899
1.4%
|
78.49
10.6%
|
0.744
11.7%
| |
Random Forest
| Morphology | 0.556 | 0.871 | 68.82 | 0.604 |
Morphology + Texture | 0.667 | 0.864 | 53.76 | 0.609 | |
ADC + SER | 0.667 | 0.9 | 70.97 | 0.764 | |
Morphology + Kinetic | 0.611 | 0.885 | 69.89 | 0.713 | |
Morphology + ADC
|
0.667
|
0.91
|
78.49
|
0.8
| |
Morphology + Texture + Kinetic | 0.667 | 0.906 | 75.27 | 0.722 | |
Entire
*%
|
0.722
8.3%
|
0.912
1%
|
69.89
-7.2%
|
0.787
9%
| |
Average
| Morphology | 0.445 | 0.846 | 67.03 | 0.566 |
Morphology + Texture | 0.518 | 0.853 | 63.44 | 0.61 | |
ADC + SER | 0.703 | 0.914 | 75.27 | 0.776 | |
Morphology + Kinetic | 0.556 | 0.876 | 71.33 | 0.681 | |
Morphology + ADC
|
0.630
|
0.873
|
78.13
|
0.778
| |
Morphology + Texture + Kinetic | 0.611 | 0.892 | 73.84 | 0.689 | |
Entire
*%
|
0.685
12.1%
|
0.912
2.2%
|
75.98
2.9%
|
0.766
11.2%
|
Scenario 2: ADC is highly diagnostic and can increase sensitivity when combined with other features
Scenario 3: Carefully selected features achieved the best diagnostic performance
Selected feature | Criteria | Specificity | Sensitivity | Accuracy | AUC |
---|---|---|---|---|---|
ADC | 0.778 | 0.94 | 82.8 | 0.809 | |
Sum average | KNN [34] | 0.667 | 0.91 | 78.50 | 0.815 |
Entropy | Random Forest [35] | 0.722 | 0.92 | 74.19 | 0.791 |
Elongation | |||||
Sum variance |
Average
|
0.722
|
0.923
|
78.50
|
0.805
|