World J Oncol

World Journal of Oncology, ISSN 1920-4531 print, 1920-454X online, Open Access

Article copyright, the authors; Journal compilation copyright, World J Oncol and Elmer Press Inc

Journal website https://www.wjon.org

Original Article

Volume 15, Number 1, February 2024, pages 58-71

Radiomics of Preoperative Multi-Sequence Magnetic Resonance Imaging Can Improve the Predictive Performance of Microvascular Invasion in Hepatocellular Carcinoma

Figure 1. The workflow of patient selection for this study. MR: magnetic resonance; MVI: microvascular invasion; HCC: hepatocellular carcinoma.

Figure 2. The workflow of the radiomics analysis. ROC: receiver operating characteristic; DCA: decision curve analysis; NRI: net reclassification improvement; IDI: integrated discrimination improvement.

Figure 3. Each fold cross-validation’s AUC and mean AUC in training/validation cohort of the three models. Clinicoradiological model’s training cohort (a) and validation cohort (b). Radiomics model’s training cohort (c) and validation cohort (d). Clinicoradiomics model’s training cohort (e) and validation cohort (f). AUC: area under the receiver operating curve; ROC: receiver operating characteristic.

Figure 4. The three models’ mean ROC curves of training cohort (a) and validation cohort (b). Calibration curves (c) and DCA curves (d) of three models for predicting MVI of HCC. AUC: area under the receiver operating curve; ROC: receiver operating characteristic; DCA: decision curve analysis; MVI: microvascular invasion; HCC: hepatocellular carcinoma.

**Table 1.** Baseline Characteristics for Predicting MVI
Characteristic	Total (n = 206)	MVI absent (n = 134)	MVI present (n = 72)	Univariable logistic regression
*P < 0.05. The original units of lg10 converted variables are AFP (ng/L), PIVKA-II (mAU/mL). P is the P value of univariate logistic regression analysis. MVI: microvascular invasion; OR: odd ratio; CI: confidence interval; BCLC: Barcelona Clinic Liver Cancer; HBV: hepatitis B virus; AFP: alpha-fetoprotein; PIVKA-II: protein induced by vitamin K absence or antagonist-II; CA199: carbohydrate antigen 19-9; APHE: arterial phase hyperenhancement; MRI: magnetic resonance imaging.
Clinical features
Age	55.19 ± 10.69	54.69 ± 10.85	56.13 ± 10.38	1.013 (0.986, 1.041)	0.357
Gender
Male	172 (83.5%)	108 (52.4%)	64 (31.1%)
Female	34 (16.5%)	26 (12.6%)	8 (3.9%)	0.519 (0.222, 1.216)	0.131
BCLC stage					0.080
0	52 (25.2%)	32 (15.5%)	20 (9.7%)
A	139 (67.5%)	96 (46.6%)	43 (20.9%)	0.717 (0.369, 1.393)	0.326
B	15 (7.3%)	6 (2.9%)	9 (4.4%)	2.400 (0.742, 7.767)	0.144
Child-Pugh stage
A	198 (96.1%)	127 (61.7%)	71 (34.5%)
B	8 (3.9%)	7 (3.4%)	1 (0.5%)	0.256 (0.031, 2.119)	0.206
Liver disease
HBV	181 (87.9%)	116 (56.3%)	65 (31.6%)
None or other	25 (12.1%)	18 (8.7%)	7 (3.4%)	0.694 (0.275, 1.749)	0.439
AFP-L3
Negative	131 (63.7%)	87 (42.2%)	44 (21.4%)
Positive	75 (36.4%)	47 (22.8%)	28 (13.6%)	1.178 (0.652, 2.129)	0.588
AFP_lg10	1.47 (0.66 - 2.34)	1.18 (0.52 - 2.13)	1.8 (0.99 - 2.65)	1.607 (1.175, 2.196)	0.003*
PIVKA-II_lg10	2.05 (1.52 - 2.79)	2.03 (1.44 - 2.72)	2.17 (1.67 - 2.98)	1.412 (1.004, 1.986)	0.047*
CA199 (U/mL)	16.45 (8.5 - 28.3)	17.3 (9.05 - 28.25)	15.1 (8.05 - 28.5)	0.992 (0.976, 1.008)	0.306
HBsAg
Negative	32 (15.5%)	22 (10.7%)	10 (4.9%)
Positive	174 (84.5%)	112 (54.4%)	62 (30.1%)	1.218 (0.542, 2.736)	0.633
Pathologic factors
Microscopic cirrhosis
Absent	128 (62.1%)	83 (40.3%)	45 (21.8%)
Present	78 (37.9%)	51 (24.8%)	27 (13.1%)	0.976 (0.541, 1.763)	0.976
Satellite nodules
Absent	183 (88.8%)	124 (60.2%)	59 (28.6%)
Present	23 (11.2%)	10 (4.9%)	13 (6.3%)	2.732 (1.132, 6.592)	0.025*
Edmondson-Steiner grade
I -II	26 (12.6%)	20 (9.7%)	6 (2.9%)
III - IV	180 (87.4%)	114 (55.3%)	66 (32%)	1.930 (0.738, 5.047)	0.180
MRI features
Tumor diameter (cm)	3 (2.18 - 4.43)	2.9 (2.2 - 4.13)	3.45 (2.03 - 5.28)	1.105 (0.945, 1.293)	0.212
Tumor number
Solitary	188 (91.3%)	126 (61.2%)	62 (30.1%)
Multiple	18 (8.7%)	8 (3.9%)	10 (4.9%)	2.540 (0.955, 6.756)	0.062
Shape
Regular	125 (60.7%)	89 (43.2%)	36 (17.5%)
Irregular	81 (39.3%)	45 (21.8%)	36 (17.5%)	1.978 (1.102, 3.549)	0.022*
Margin
Smooth	103 (50%)	78 (37.9%)	25 (12.1%)
Non-smooth	103 (50%)	56 (27.2%)	47 (22.8%)	2.619 (1.445, 4.744)	0.002*
Radiological capsule enhancement					< 0.001*
Complete	79 (38.3%)	68 (33%)	11 (5.3%)
Incomplete	86 (41.7%)	44 (21.4%)	42 (20.4%)	5.901 (2.747, 12.675)	< 0.001*
Absent	41 (19.9%)	22 (10.7%)	19 (9.2%)	5.339 (2.204, 12.931)	< 0.001*
Restricted diffusion
Present	193 (93.7%)	123 (59.7%)	70 (34%)
Absent	13 (6.3%)	11 (5.3%)	2 (1%)	0.319 (0.069, 1.483)	0.145
Nonrim APHE
Present	123 (59.7%)	97 (47.1%)	26 (12.6%)
Absent	83 (40.3%)	37 (18%)	46 (22.3%)	4.638 (2.515, 8.554)	< 0.001*
Rim APHE
Absent	136 (66%)	104 (50.5%)	32 (15.5%)
Present	70 (34%)	30 (14.6%)	40 (19.4%)	4.333 (2.337, 8.034)	< 0.001*
Non-peripheral “washout”
Present	123 (59.7%)	92 (44.7%)	31 (15 %)
Absent	83 (40.3%)	42 (20.4%)	41 (19.9%)	2.897 (1.602, 5.238)	< 0.001*
Enhancement pattern
Typical	118 (57.3%)	91 (44.2%)	27 (13.1%)
Atypical	88 (42.7%)	43 (20.9%)	45 (21.8%)	3.527 (1.937, 6.423)	< 0.001*
Arterial peritumoral enhancement
Absent	134 (65%)	109 (52.9%)	25 (12.1%)
Present	72 (35%)	25 (12.1%)	47 (22.8%)	8.197 (4.273, 15.723)	< 0.001*
MRI liver cirrhosis
Absent	57 (27.7%)	44 (21.4%)	13 (6.3%)
Present	149 (72.3%)	90 (43.7%)	59 (28.6%)	2.219 (1.101, 4.470)	0.026*

Table 1. Baseline Characteristics for Predicting MVI

Characteristic

Total (n = 206)

MVI absent (n = 134)

MVI present (n = 72)

Univariable logistic regression

OR (95%CI)

*P < 0.05. The original units of lg10 converted variables are AFP (ng/L), PIVKA-II (mAU/mL). P is the P value of univariate logistic regression analysis. MVI: microvascular invasion; OR: odd ratio; CI: confidence interval; BCLC: Barcelona Clinic Liver Cancer; HBV: hepatitis B virus; AFP: alpha-fetoprotein; PIVKA-II: protein induced by vitamin K absence or antagonist-II; CA199: carbohydrate antigen 19-9; APHE: arterial phase hyperenhancement; MRI: magnetic resonance imaging.

Clinical features

Age

55.19 ± 10.69

54.69 ± 10.85

56.13 ± 10.38

1.013 (0.986, 1.041)

0.357

Gender

Male

172 (83.5%)

108 (52.4%)

64 (31.1%)

Female

34 (16.5%)

26 (12.6%)

8 (3.9%)

0.519 (0.222, 1.216)

0.131

BCLC stage

0.080

52 (25.2%)

32 (15.5%)

20 (9.7%)

139 (67.5%)

96 (46.6%)

43 (20.9%)

0.717 (0.369, 1.393)

0.326

15 (7.3%)

6 (2.9%)

9 (4.4%)

2.400 (0.742, 7.767)

0.144

Child-Pugh stage

198 (96.1%)

127 (61.7%)

71 (34.5%)

8 (3.9%)

7 (3.4%)

1 (0.5%)

0.256 (0.031, 2.119)

0.206

Liver disease

HBV

181 (87.9%)

116 (56.3%)

65 (31.6%)

None or other

25 (12.1%)

18 (8.7%)

7 (3.4%)

0.694 (0.275, 1.749)

0.439

AFP-L3

Negative

131 (63.7%)

87 (42.2%)

44 (21.4%)

Positive

75 (36.4%)

47 (22.8%)

28 (13.6%)

1.178 (0.652, 2.129)

0.588

AFP_lg10

1.47 (0.66 - 2.34)

1.18 (0.52 - 2.13)

1.8 (0.99 - 2.65)

1.607 (1.175, 2.196)

0.003*

PIVKA-II_lg10

2.05 (1.52 - 2.79)

2.03 (1.44 - 2.72)

2.17 (1.67 - 2.98)

1.412 (1.004, 1.986)

0.047*

CA199 (U/mL)

16.45 (8.5 - 28.3)

17.3 (9.05 - 28.25)

15.1 (8.05 - 28.5)

0.992 (0.976, 1.008)

0.306

HBsAg

Negative

32 (15.5%)

22 (10.7%)

10 (4.9%)

Positive

174 (84.5%)

112 (54.4%)

62 (30.1%)

1.218 (0.542, 2.736)

0.633

Pathologic factors

Microscopic cirrhosis

Absent

128 (62.1%)

83 (40.3%)

45 (21.8%)

Present

78 (37.9%)

51 (24.8%)

27 (13.1%)

0.976 (0.541, 1.763)

0.976

Satellite nodules

Absent

183 (88.8%)

124 (60.2%)

59 (28.6%)

Present

23 (11.2%)

10 (4.9%)

13 (6.3%)

2.732 (1.132, 6.592)

0.025*

Edmondson-Steiner grade

I -II

26 (12.6%)

20 (9.7%)

6 (2.9%)

III - IV

180 (87.4%)

114 (55.3%)

66 (32%)

1.930 (0.738, 5.047)

0.180

MRI features

Tumor diameter (cm)

3 (2.18 - 4.43)

2.9 (2.2 - 4.13)

3.45 (2.03 - 5.28)

1.105 (0.945, 1.293)

0.212

Tumor number

Solitary

188 (91.3%)

126 (61.2%)

62 (30.1%)

Multiple

18 (8.7%)

8 (3.9%)

10 (4.9%)

2.540 (0.955, 6.756)

0.062

Shape

Regular

125 (60.7%)

89 (43.2%)

36 (17.5%)

Irregular

81 (39.3%)

45 (21.8%)

36 (17.5%)

1.978 (1.102, 3.549)

0.022*

Margin

Smooth

103 (50%)

78 (37.9%)

25 (12.1%)

Non-smooth

103 (50%)

56 (27.2%)

47 (22.8%)

2.619 (1.445, 4.744)

0.002*

Radiological capsule enhancement

< 0.001*

Complete

79 (38.3%)

68 (33%)

11 (5.3%)

Incomplete

86 (41.7%)

44 (21.4%)

42 (20.4%)

5.901 (2.747, 12.675)

< 0.001*

Absent

41 (19.9%)

22 (10.7%)

19 (9.2%)

5.339 (2.204, 12.931)

< 0.001*

Restricted diffusion

Present

193 (93.7%)

123 (59.7%)

70 (34%)

Absent

13 (6.3%)

11 (5.3%)

2 (1%)

0.319 (0.069, 1.483)

0.145

Nonrim APHE

Present

123 (59.7%)

97 (47.1%)

26 (12.6%)

Absent

83 (40.3%)

37 (18%)

46 (22.3%)

4.638 (2.515, 8.554)

< 0.001*

Rim APHE

Absent

136 (66%)

104 (50.5%)

32 (15.5%)

Present

70 (34%)

30 (14.6%)

40 (19.4%)

4.333 (2.337, 8.034)

< 0.001*

Non-peripheral “washout”

Present

123 (59.7%)

92 (44.7%)

31 (15 %)

Absent

83 (40.3%)

42 (20.4%)

41 (19.9%)

2.897 (1.602, 5.238)

< 0.001*

Enhancement pattern

Typical

118 (57.3%)

91 (44.2%)

27 (13.1%)

Atypical

88 (42.7%)

43 (20.9%)

45 (21.8%)

3.527 (1.937, 6.423)

< 0.001*

Arterial peritumoral enhancement

Absent

134 (65%)

109 (52.9%)

25 (12.1%)

Present

72 (35%)

25 (12.1%)

47 (22.8%)

8.197 (4.273, 15.723)

< 0.001*

MRI liver cirrhosis

Absent

57 (27.7%)

44 (21.4%)

13 (6.3%)

Present

149 (72.3%)

90 (43.7%)

59 (28.6%)

2.219 (1.101, 4.470)

0.026*

**Table 2.** Univariate and Multivariate Logistic Regression Analysis for Predicting MVI
Variable	Univariable logistic regression	Multivariable logistic regression
P is the P value of univariate and multivariate logistic regression analysis. MVI: microvascular invasion; OR: odd ratio; CI: confidence interval; AFP: alpha-fetoprotein; PIVKA-II: protein induced by vitamin K absence or antagonist-II; APHE: arterial phase hyperenhancement; MRI: magnetic resonance imaging.
AFP_lg10	1.607	1.175, 2.196	0.003	1.469	1.002, 2.154	0.049
PIVKA-II_lg10	1.412	1.004, 1.986	0.047
Shape (irregular)	1.978	1.102, 3.549	0.022
Margin (non-smooth)	2.619	1.445, 4.744	0.002
Radiological capsule enhancement			< 0.001			0.002
Incomplete	5.901	2.747, 12.675	< 0.001	4.101	1.721, 9.770	0.001
Absent	5.339	2.204, 12.931	< 0.001	5.193	1.797, 15.009	0.002
Nonrim APHE (absent)	4.638	2.515, 8.554	< 0.001
Rim APHE (present)	4.333	2.337, 8.034	< 0.001
Non-peripheral “washout” (absent)	2.897	1.602, 5.238	< 0.001
Enhancement pattern (atypical)	3.527	1.937, 6.423	< 0.001	2.793	1.358, 5.742	0.005
Arterial peritumoral enhancement (present)	8.197	4.273, 15.723	< 0.001	8.222	3.917, 17.259	< 0.001
MRI liver cirrhosis (present)	2.219	1.101, 4.470	0.026

Table 2. Univariate and Multivariate Logistic Regression Analysis for Predicting MVI

Variable

Univariable logistic regression

Multivariable logistic regression

95% CI

P is the P value of univariate and multivariate logistic regression analysis. MVI: microvascular invasion; OR: odd ratio; CI: confidence interval; AFP: alpha-fetoprotein; PIVKA-II: protein induced by vitamin K absence or antagonist-II; APHE: arterial phase hyperenhancement; MRI: magnetic resonance imaging.

AFP_lg10

1.607

1.175, 2.196

0.003

1.469

1.002, 2.154

0.049

PIVKA-II_lg10

1.412

1.004, 1.986

0.047

Shape (irregular)

1.978

1.102, 3.549

0.022

Margin (non-smooth)

2.619

1.445, 4.744

0.002

Radiological capsule enhancement

< 0.001

0.002

Incomplete

5.901

2.747, 12.675

< 0.001

4.101

1.721, 9.770

0.001

Absent

5.339

2.204, 12.931

< 0.001

5.193

1.797, 15.009

0.002

Nonrim APHE (absent)

4.638

2.515, 8.554

< 0.001

Rim APHE (present)

4.333

2.337, 8.034

< 0.001

Non-peripheral “washout” (absent)

2.897

1.602, 5.238

< 0.001

Enhancement pattern (atypical)

3.527

1.937, 6.423

< 0.001

2.793

1.358, 5.742

0.005

Arterial peritumoral enhancement (present)

8.197

4.273, 15.723

< 0.001

8.222

3.917, 17.259

< 0.001

MRI liver cirrhosis (present)

2.219

1.101, 4.470

0.026

**Table 3.** The Mean Performances of Diverse Sequences After Stratified 5-Fold Cross-Validation for Predicting MVI
Sequences (feature number)/model	AUC	f1_score	Accuracy	Sensitivity	Specificity
The training and validation cohort’s performances of each fold for predicting MVI was shown here (Supplementary Material 5, www.wjon.org). ALL = (AP + PP + DP + DWI + T1 + T2), was also the radiomics model. MVI: microvascular invasion; AUC: area under receiver operating characteristic curve; T2: T2-weighted imaging; T1: T1-weighted imaging; DWI: diffusion-weighted imaging; AP: arterial phase; PP: portal phase; DP: delayed phase.
AP (20)	0.842	0.805	0.721	0.624	0.825	0.767	0.667	0.610	0.848	0.838
PP (15)	0.831	0.808	0.635	0.539	0.778	0.718	0.839	0.817	0.697	0.697
DP (13)	0.808	0.777	0.634	0.542	0.783	0.738	0.688	0.653	0.808	0.808
DWI (14)	0.827	0.802	0.634	0.526	0.609	0.743	0.827	0.750	0.697	0.778
T2 (13)	0.772	0.736	0.612	0.429	0.584	0.694	0.693	0.734	0.727	0.667
T1 (16)	0.810	0.770	0.626	0.596	0.647	0.576	0.683	0.610	0.778	0.798
AP + DWI (16)	0.870	0.855	0.711	0.671	0.818	0.786	0.708	0.634	0.848	0.919
AP + PP (17)	0.855	0.819	0.746	0.616	0.842	0.762	0.673	0.628	0.889	0.889
AP + DP (15)	0.845	0.823	0.680	0.639	0.805	0.776	0.721	0.749	0.798	0.737
PP + DP (17)	0.854	0.810	0.719	0.618	0.824	0.772	0.702	0.791	0.838	0.667
PP + DWI (18)	0.872	0.855	0.720	0.659	0.822	0.787	0.770	0.807	0.848	0.798
T1 + T2 (20)	0.857	0.824	0.726	0.626	0.830	0.762	0.813	0.903	0.737	0.667
AP + PP + DP (15)	0.859	0.835	0.718	0.669	0.822	0.791	0.727	0.791	0.838	0.737
AP + PP + DWI (19)	0.889	0.860	0.783	0.656	0.858	0.777	0.857	0.834	0.778	0.768
AP + DP + DWI (20)	0.916	0.897	0.809	0.788	0.876	0.859	0.803	0.790	0.889	0.889
PP + DP + DWI (20)	0.906	0.878	0.810	0.712	0.881	0.815	0.813	0.831	0.848	0.778
AP + PP + DP + DWI (20)	0.920	0.904	0.829	0.707	0.888	0.811	0.824	0.790	0.889	0.879
ALL/radiomics (18)	0.925	0.907	0.823	0.722	0.883	0.820	0.844	0.820	0.848	0.848
Clinicoradiologic	0.849	0.846	0.643	0.576	0.779	0.743	0.843	0.930	0.727	0.667
Clinicoradiomics	0.950	0.933	0.887	0.812	0.779	0.869	0.879	0.890	0.889	0.838

Table 3. The Mean Performances of Diverse Sequences After Stratified 5-Fold Cross-Validation for Predicting MVI

Sequences (feature number)/model

AUC

f1_score

Accuracy

Sensitivity

Specificity

Training

Validation

Training

Validation

Training

Validation

Training

Validation

Training

Validation

The training and validation cohort’s performances of each fold for predicting MVI was shown here (Supplementary Material 5, www.wjon.org). ALL = (AP + PP + DP + DWI + T1 + T2), was also the radiomics model. MVI: microvascular invasion; AUC: area under receiver operating characteristic curve; T2: T2-weighted imaging; T1: T1-weighted imaging; DWI: diffusion-weighted imaging; AP: arterial phase; PP: portal phase; DP: delayed phase.

AP (20)

0.842

0.805

0.721

0.624

0.825

0.767

0.667

0.610

0.848

0.838

PP (15)

0.831

0.808

0.635

0.539

0.778

0.718

0.839

0.817

0.697

DP (13)

0.808

0.777

0.634

0.542

0.783

0.738

0.688

0.653

0.808

DWI (14)

0.827

0.802

0.634

0.526

0.609

0.743

0.827

0.750

0.697

0.778

T2 (13)

0.772

0.736

0.612

0.429

0.584

0.694

0.693

0.734

0.727

0.667

T1 (16)

0.810

0.770

0.626

0.596

0.647

0.576

0.683

0.610

0.778

0.798

AP + DWI (16)

0.870

0.855

0.711

0.671

0.818

0.786

0.708

0.634

0.848

0.919

AP + PP (17)

0.855

0.819

0.746

0.616

0.842

0.762

0.673

0.628

0.889

AP + DP (15)

0.845

0.823

0.680

0.639

0.805

0.776

0.721

0.749

0.798

0.737

PP + DP (17)

0.854

0.810

0.719

0.618

0.824

0.772

0.702

0.791

0.838

0.667

PP + DWI (18)

0.872

0.855

0.720

0.659

0.822

0.787

0.770

0.807

0.848

0.798

T1 + T2 (20)

0.857

0.824

0.726

0.626

0.830

0.762

0.813

0.903

0.737

0.667

AP + PP + DP (15)

0.859

0.835

0.718

0.669

0.822

0.791

0.727

0.791

0.838

0.737

AP + PP + DWI (19)

0.889

0.860

0.783

0.656

0.858

0.777

0.857

0.834

0.778

0.768

AP + DP + DWI (20)

0.916

0.897

0.809

0.788

0.876

0.859

0.803

0.790

0.889

PP + DP + DWI (20)

0.906

0.878

0.810

0.712

0.881

0.815

0.813

0.831

0.848

0.778

AP + PP + DP + DWI (20)

0.920

0.904

0.829

0.707

0.888

0.811

0.824

0.790

0.889

0.879

ALL/radiomics (18)

0.925

0.907

0.823

0.722

0.883

0.820

0.844

0.820

0.848

Clinicoradiologic

0.849

0.846

0.643

0.576

0.779

0.743

0.843

0.930

0.727

0.667

Clinicoradiomics

0.950

0.933

0.887

0.812

0.779

0.869

0.879

0.890

0.889

0.838

**Table 4.** Comparison Between Models
	P (AUC)	NRI	P (NRI)	IDI	P (IDI)
P is the probability when Delong test is used to compare the ROC curve of two models. AUC: area under receiver operating characteristic curve; NRI: net reclassification improvement; IDI: integrated discrimination improvement.
Radiomics vs. clinicoradiologic	0.0017	0.575	0.0014	0.280	< 0.05
Clinicoradiomics vs. clinicoradiologic	< 0.0001	0.825	< 0.0001	0.398	< 0.05
Clinicoradiomics vs. radiomics	0.0122	0.313	0.0085	0.117	< 0.05

Table 4. Comparison Between Models

P (AUC)

NRI

P (NRI)

IDI

P (IDI)

P is the probability when Delong test is used to compare the ROC curve of two models. AUC: area under receiver operating characteristic curve; NRI: net reclassification improvement; IDI: integrated discrimination improvement.

Radiomics vs. clinicoradiologic

0.0017

0.575

0.0014

0.280

< 0.05

Clinicoradiomics vs. clinicoradiologic

< 0.0001

0.825

< 0.0001

0.398

< 0.05

Clinicoradiomics vs. radiomics

0.0122

0.313

0.0085

0.117

< 0.05