Patients’ Characteristics
A total of 497 female patients with 548 pathologically confirmed breast lesions (274 benign and 274 malignant) assessed as BI-RADS 3, 4, or 5 lesions on preoperative breast MRI examinations from June 2018 to September 2021 were enrolled in this retrospective study.
Of the 344 lesions in the training cohort, 33 were BI-RADS 3 (benign), 298 were BI-RADS 4 (155/298 [52.0%] malignant), and 13 were BI-RADS 5 (12/13 [92.3%] malignant). Of the 204 lesions in the validation cohort, 21 were BI-RADS 3 (all were benign), 170 were BI-RADS 4 (94/140 [67.1%] malignant), and 13 were BI-RADS 5 (all were malignant). The demographic and histopathological data of the patients in the training and validation cohorts are displayed in Table 1.
Table 1. Baseline Characteristics and Histopathologic Results of Lesions in the Training and Validation Cohorts
Characteristic
|
Training set
(n = 344)
|
|
Validation cohort
(n = 204)
|
P value
|
Age (years)
|
43.8 ± 12.0
|
|
47.9 ± 11.2
|
0.65
|
Diameter (cm)
|
1.5 [0.6–11.4]
|
|
1.6 [0.6–6.2]
|
0.20
|
Lesion subsets
|
|
|
|
0.74
|
Mass
|
266 (77.4%)
|
|
161 (78.9%)
|
|
NME
|
78 (22.6%)
|
|
43 (21.1%)
|
|
Lesion diameter
|
|
|
|
0.03
|
Diameter ≤1 cm
|
97 (28.1%)
|
|
40 (19.6%)
|
|
Diameter >1 cm
|
247 (71.9%)
|
|
164 (80.4%)
|
|
Number of BI-RADS 3, 4, or 5 lesions
|
|
|
|
0.27
|
BI-RADS 3
|
33 (9.6%)
|
|
21 (10.3%)
|
|
BI-RADS 4
|
298 (86.6%)
|
|
170 (83.3%)
|
|
BI-RADS 5
|
13 (3.8%)
|
|
13 (6.4%0
|
|
Table 1. Continued
Characteristic
|
Training set
(n = 344)
|
|
Validation cohort
(n = 204)
|
P value
|
Pathologic results of malignancy
|
|
|
|
0.46
|
IDC
|
123 (35.8%)
|
|
79 (38.7%)
|
|
DCIS
|
16 (4.7%)
|
|
17 (8.3%)
|
|
Mucoid carcinoma
|
7 (2.0%)
|
|
2 (1.0%)
|
|
IPC
|
9 (2.6%)
|
|
4 (2.0%)
|
|
ILC
|
5 (1.5%)
|
|
3 (1.5%)
|
|
Other malignant lesions
|
7 (2.0%)
|
|
2 (1.0%)
|
|
Pathologic results of malignancy
|
|
|
|
0.88
|
Fibroadenoma
|
86 (25.0%)
|
|
52 (25.5%)
|
|
Adenosis
|
39 (11.3%)
|
|
20 (9.8%)
|
|
Papilloma
|
42 (12.2%)
|
|
17 (7.1%)
|
|
Mastitis
|
7 (2.0%)
|
|
4 (2.0%)
|
|
Apocrine hidradenoma
|
1 (0.3%)
|
|
0 (0%)
|
|
Myoepithelioma
|
1 (0.3%)
|
|
1 (0.5%)
|
|
Phyllode tumour
|
1 (0.3%)
|
|
1 (0.5%)
|
|
|
|
|
|
|
|
Data are presented as mean ± standard deviation, median (interquartile range), or number percentage.
NME non-mass enhancement, BI-RADS Breast Imaging Reporting and Data System, DCIS ductal carcinoma in situ, IDC invasive ductal carcinoma, ILC invasive lobular carcinoma, IPC intraductal papillary carcinoma.
Diagnostic Performance of the ADCmin and ADCmean
Among all subset analyses, the ADCmin and ADCmean of the malignant lesions were significantly lower than those of the benign lesions (P < 0.05). No significant difference was noted between ADCmin and ADCmean in terms of their diagnostic performance according to the Hanley and McNeil test (P > 0.05). Moreover, the ADCmean and ADCmin were worse in non-mass lesions than in mass lesions. When stratified according to size, the ADCmin and ADCmean performed worse in lesions ≤ 1 cm than in lesions >1 cm, particularly in the validation cohort. The ADCmin and ADCmean analysis results for all the lesion subsets are displayed in Table 2.
Diagnostic Performance of the ADCnl-min and ADCnl-mean
In the training cohort, when the -LR was 0.1, the ADCmin and ADCmean were 1.39×10−3 mm2/s (ADCnl-min) and 1.52 × 10−3 mm2/s (ADCnl-mean), respectively, with the same sensitivity of 0.946 [CI: 0.900–0.975]. The negative predictive value (NPV) was 0.907 [CI: 0.867–0.954] and 0.903 [CI: 0.863–0.953], respectively. Regarding the validation cohort, the ADCnl-min and ADCnl-mean showed good performance in distinguishing between benign and malignant breast lesions with an -LR of ≤ 0.1. The sensitivity was 0.963 [CI: 0.907–0.990] and 0.944 [CI: 0.882–0.979], and the NPV was 0.938 [CI: 0.875–0.981] and 0.901 [CI: 0.855–0.961], respectively. The corresponding diagnostic parameters of the ADCnl-min and ADCnl-mean for the validation cohort are displayed in Table 3. The -LR of 3/5 subsets was ≤ 0.1 in the training and validation cohorts.
Table 2. ADC and AUCs by Lesion Subsets
ADC
|
Training cohort
|
|
Validation cohort
|
AUC
|
95%CI
|
P-value
|
|
AUC
|
95%CI
|
P-value
|
All lesions
|
|
|
|
|
|
|
|
ADCmin
|
0.873
|
0.828–0.902
|
Reference
|
|
0.863
|
0.808–0.907
|
Reference
|
ADCmean
|
0.868
|
0.833–0.907
|
0.533
|
|
0.881
|
0.829–0.922
|
0.092
|
Mass
|
|
|
|
|
|
|
|
ADCmin
|
0.896
|
0.853–0.930
|
Reference
|
|
0.888
|
0.828–0.932
|
Reference
|
ADCmean
|
0.894
|
0.850–0.928
|
0.736
|
|
0.901
|
0.854–0.949
|
0.052
|
NME
|
|
|
|
|
|
|
|
ADCmin
|
0.774
|
0.666–0.861
|
Reference
|
|
0.755
|
0.600–0.873
|
Reference
|
ADCmean
|
0.761
|
0.651–0.850
|
0.414
|
|
0.767
|
0.613–0.882
|
0.733
|
Diameter ≤1 cm
|
|
|
|
|
|
|
|
ADCmin
|
0.867
|
0.782–0.927
|
Reference
|
|
0.693
|
0.528–0.829
|
Reference
|
ADCmean
|
0.855
|
0.769–0.918
|
0.593
|
|
0.747
|
0.584–0.873
|
0.145
|
Table 2. Continued
ADC
|
Training cohort
|
|
Validation cohort
|
AUC
|
95%CI
|
P-value
|
|
AUC
|
95% CI
|
P-value
|
Diameter >1 cm
|
|
|
|
|
|
|
|
ADCmin
|
0.878
|
0.830–0.916
|
Reference
|
|
0.908
|
0.853–0.947
|
Reference
|
ADCmean
|
0.873
|
0.824–0.911
|
0.573
|
|
0.918
|
0.865–0.955
|
0.291
|
ADCmin minimum apparent diffusion coefficient, ADCmean mean apparent diffusion coefficient, AUC area under the receiver operating characteristic curve, 95% CI 95% confidence interval, NME non-mass enhancement ADC apparent diffusion coefficient.
Table 3. Performance of the ADCnl-min and ADCnl-mean
|
AUC (95%CI)
|
-LR (95%CI)
|
Sen (95%CI)
|
Spe (95%CI)
|
NPV (95%CI)
|
PPV (95%CI)
|
All lesions
|
|
|
|
|
|
|
ADCnl-min
|
0.796
(0.734–0.849)
|
0.06
(0.02–0.2)
|
0.963
(0.907–0.990)
|
0.629
(0.525–0.725)
|
0.938
(0.875–0.981)
|
0.763
(0.718–0.805)
|
ADCnl-mean
|
0.781
(0.718–0.836)
|
0.09
(0.04–0.2)
|
0.944
(0.882–0.979)
|
0.619
(50.4 –70.6%)
|
0.901
(0.855–0.961)
|
0.732
(0.685–0.768)
|
Mass
|
|
|
|
|
|
|
ADCnl-min
|
0.811
(0.742–0.868)
|
0.02
(0.003–0.1)
|
0.987
(0.931–1.00)
|
0.634
(0.520–0.738)
|
0.981
(0.917–1.0)
|
0.722
(0.669–0.779)
|
ADCnl-mean
|
0.811
(0.742–0.868)
|
0.02
(0.003–0.1)
|
0.987
(0.931–1.00)
|
0.634
(0.520–0.738)
|
0.981
(0.917–1.0)
|
0.722
(0.669–0.779)
|
NME
|
|
|
|
|
|
|
ADCnl-min
|
0.746
(0.591–0.867)
|
0.18
(0.06–0.6)
|
0.893
(0.718–0.977)
|
0.600
(0.323–0.837)
|
0.750
(0.619–0.833)
|
0.781
(0.730–0.909)
|
ADCnl-mean
|
0.677
(0.518–0.812)
|
0.33
(0.1–0.8)
|
0.821
(0.631–0.939)
|
0.533
(0.266–0.787)
|
0.615
(0.534–0.702)
|
0.767
(0.705–0.828)
|
Diameter ≤1 cm
|
|
|
|
|
|
|
ADCnl-min
|
0.747
(0.584–0.871)
|
0.12
(0.02–0.8)
|
0.933
(0.681–0.998)
|
0.560
(0.349–0.766)
|
0.933
(0.791–0.998)
|
0.560
(0.481–0.625)
|
ADCnl-mean
|
0.727
(0.563–0.855)
|
0.11
(0.02–0.8)
|
0.933
(0.681–0.998)
|
0.520
(0.31–0.722)
|
0.929
(0.790–1.0)
|
0.538
(0.466–0.595)
|
Diameter >1 cm
|
|
|
|
|
|
|
ADCnl-min
|
0.810
(0.742–0.867)
|
0.05
(0.02–0.2)
|
0.967
(0.908–0.993)
|
0.653
(0.531–0.761)
|
0.940
(0.873–0.974)
|
0.781
(0.728–0.831)
|
ADCnl-mean
|
0.799
(0.730–0.858)
|
0.08
(0.03–0.2)
|
0.964
(0.878–0.982)
|
0.639
(0.529–0.740)
|
0.904
(0.826–0.957)
|
0.780
(0.727–0.812)
|
ADCnl-min the value of minimum apparent diffusion coefficient with the highest specificity when negative likelihood ratios are ≤ 0.1, ADCnl-mean the value of mean apparent diffusion coefficient with the highest specificity when negative likelihood ratios are ≤ 0.1, AUC area under the receiver operating characteristic curve, -LR negative likelihood ratio, 95% CI 95% confidence interval, Sen sensitivity, PPV positive predictive value, NPV negative predictive value, NME non-mass enhancement.
Comparison of the Biopsy-Reducing Ability
Hypothetically, lesions with ADCmin >ADCnl-min or ADCmean >ADCnl-mean were considered benign and not biopsied. For the BI-RADS 4 and 5 lesions, unnecessary biopsy rates were reduced by 25.1% [CI: 0.196–0.312] and 24.1% [CI: 0.186–0.305], respectively, in the training cohort. In the validation cohort, unnecessary biopsy rates were reduced by 26.2% [CI: 0.202–0.339] and 26.7% [CI: 0.197–0.351], respectively. There was no significant difference among all subsets in the reduction of the biopsy rate of BI-RADS 4 and 5 lesions between the ADCnl-min and ADCnl-mean (P > 0.05) (Figure 4).
Comparison of the PPV and False-Negative Number
Compared with the BI-RADS assessment, after applying the ADCnl-min and ADCnl-mean, both significantly differed from BI-RADS regarding the PPV. The application of the ADCnl-min and ADCnl-mean resulted in an increase of 14.1% and 13.5% in the PPV, respectively, in the training cohort, and an increase of 17.8% and 14.7% in the PPV, respectively, in the test cohort. No significant difference was noted between the ADCnl-min and ADCnl-mean regarding the PPV (Table 4).
Table 4. Improved PPV in the training and validation cohorts
Method
|
Training cohort
|
|
Validation cohort
|
PPV
|
P-value
|
|
PPV
|
P-value
|
BI-RADS
|
53.7%
(167/311)
|
Reference
|
Reference
|
|
58.5%
(107/183)
|
Reference
|
Reference
|
ADCnl-min
|
67.8%
(158/233)
|
< 0.001
|
Reference
|
|
76.3%
(103/135)
|
< 0.001
|
Reference
|
ADCnl-mean
|
66.9%
(158/236)
|
0.002
|
0.84
|
|
73.2%
(101/138)
|
0.009
|
0.65
|
ADCnl-min the value of minimum apparent diffusion coefficient with the highest specificity when negative likelihood ratios are ≤ 0.1, ADCnl-mean the value of mean apparent diffusion coefficient with the highest specificity when negative likelihood ratios are ≤ 0.1, PPV positive predictive value, ADC apparent diffusion coefficient, BI-RADS Breast Imaging Reporting and Data System
In the training cohort, 9 of the 177 malignant lesions would have been incorrectly classified as non-suspicious after applying the ADCnl-min and ADCnl-mean. The 9 false-negative lesions were DCIS (n = 2), mucinous carcinoma (n = 1), mucinous carcinoma with other cancers (n = 5), and intraductal papillary carcinoma (n = 1). In the validation cohort, 4 of the 107 malignant lesions and 6 of the 107 malignant lesions would have been incorrectly classified as non-suspicious after applying the ADCnl-min and ADCnl-mean. The 4 false-negative lesions were DCIS (n = 3) and mucinous carcinoma with other cancers (n = 1); the 6 false-negative lesions were DCIS (n = 5) and mucinous carcinoma with other cancers (n = 1).
Performance of the ADC Threshold with 100% Sensitivity and Maximal Specificity
In the training set, ADCmin and ADCmean with 100% sensitivity and maximal specificity were 2.20 × 10-3 mm2/s and 2.27 × 10-3 mm2/s, respectively, resulting in a 1.9% (6/311) and 2.6% (8/311) reduction in the biopsy rate for BI-RADS 4 and BI-RADS 5 lesions, respectively. Moreover, in the validation cohort, the reduction in the biopsy rates for BI-RADS 4 and BI-RADS 5 lesions was 0.5% (1/183) and 1.6% (3/183), respectively. Although the sensitivity showed slight improvement, the reduction in the biopsy rate was significantly lower than after applying the ADCnl-min and ADCnl-mean (P < 0.05).