Preoperative prediction of perineural invasion of rectal cancer based on a magnetic resonance imaging radiomics model: A dual-center study

Table 1 Magnetic resonance imaging parameters of each sequence

Scanner	Sequence	Orientation	TR (ms)	TE (ms)	FOV (mm2)	Matrix	Thickness (mm)
Siemens Aera 1.5T	T2WI	Oblique axial	4480	87	190 × 190	320 × 320	3
	T1CE	Oblique axial	7	3	210 × 210	256 × 256	3
GE Signa HDxt 1.5T	T2WI	Oblique axial	4120	70	180 × 180	256 × 192	3
	T1CE	Oblique axial	6	3	200 × 200	288 × 160	3

TR: Repetition time; TE: Echo time; FOV: Field of view; T2WI: T2-weighted imaging; T1CE: Contrast-enhanced T1WI.

Table 2 Characteristics of patients in the training, internal validation and external validation groups, n (%)

Characteristics	Training group (n = 118)		P value	Internal validation group (n = 52)		P value	External validation group (n = 74)		P value
	PNI+ (n = 39)	PNI- (n = 79)		PNI+ (n = 17)	PNI- (n = 35)		PNI+ (n = 25)	PNI- (n = 49)
Sex			0.854			0.882			0.640
Male	25 (64.1)	52 (65.8)		12 (70.6)	24 (68.6)		16 (64.0)	34 (69.4)
Female	14 (35.9)	27 (34.2)		5 (29.4)	11 (31.4)		9 (36.0)	15 (30.6)
Age (yr)	68 (56.5-73)	68 (58.5-75)	0.479	66.59 ± 12.57	64.17 ± 10.32	0.497	65.24 ± 11.70	66.84 ± 9.85	0.562
CEA			0.150			0.027			0.005
Negative (< 5 ng/mL)	23 (59.0)	57 (72.2)		9 (52.9)	30 (85.7)		14 (56.0)	42 (85.7)
Positive (≥ 5 ng/mL)	16 (41.0)	22 (27.8)		8 (47.1)	5 (14.3)		11 (44.0)	7 (14.3)
CA19-9			0.034			0.019			1.000
Negative (< 30 U/mL)	30 (76.9)	72 (91.1)		12 (70.6)	34 (97.1)		23 (92.0)	46 (93.9)
Positive (≥ 30 U/mL)	9 (23.1)	7 (8.9)		5 (29.4)	1 (2.9)		2 (8.0)	3 (6.1)
DIS			0.050			0.444			0.823
High	24 (61.5)	31 (39.2)		9 (52.9)	15 (42.9)		12 (48.0)	22 (44.9)
Mid	3 (7.7)	16 (20.3)		5 (29.4)	9 (25.7)		5 (20.0)	13 (26.5)
Low	12 (30.8)	32 (40.5)		3 (17.6)	11 (31.4)		8 (32.0)	14 (28.6)
mT			0.211			0.885			0.314
mT1-2	3 (7.7)	15 (19.0)		6 (35.3)	10 (28.6)		5 (20.0)	9 (18.4)
mT3	24 (61.5)	47 (59.5)		8 (47.1)	18 (51.4)		14 (56.0)	32 (65.3)
mT4	12 (30.8)	17 (21.5)		3 (17.6)	7 (20.0)		6 (24.0)	8 (16.3)
mN			0.003			0.063			0.028
mN0	5 (12.8)	27 (34.2)		3 (17.6)	18 (51.4)		6 (24.0)	19 (38.8)
mN1	13 (33.3)	33 (41.8)		5 (29.4)	7 (20.0)		7 (28.0)	21 (42.9)
mN2	21 (53.8)	19 (24.1)		9 (52.9)	10 (28.6)		12 (48.0)	9 (18.4)
cM			0.070			1.000			0.064
cM0	34 (87.2)	77 (97.5)		17 (100.0)	34 (97.1)		22 (88.0)	49 (100.0)
cM1	5 (12.8)	2 (2.5)		0 (0.0)	1 (2.9)		3 (12.0)	0 (0.0)
cTNM			1.000			1.000			1.000
Ⅰ	1 (2.6)	29 (36.7)		0 (0.0)	16 (45.7)		0 (0.0)	18 (36.7)
Ⅱ	9 (23.1)	25 (31.6)		5 (29.4)	13 (37.1)		3 (12.0)	12 (24.5)
Ⅲ	25 (64.1)	23 (29.1)		11 (64.7)	5 (14.3)		19 (76.0)	19 (38.8)
Ⅳ	4 (10.3)	2 (2.5)		1 (5.9)	1 (2.9)		3 (12.0)	0 (0.0)
mCRM			0.002			0.374			0.236
Negative	14 (35.9)	52 (65.8)		9 (52.9)	23 (65.7)		15 (60.0)	36 (73.5)
Positive	25 (64.1)	27 (34.2)		8 (47.1)	12 (34.3)		10 (40.0)	13 (26.5)
mEMVI			0.016			0.935			0.009
Negative	15 (38.5)	49 (62.0)		10 (58.8)	21 (60.0)		10 (40.0)	35 (71.4)
Positive	24 (61.5)	30 (38.0)		7 (41.2)	14 (40.0)		15 (60.0)	14 (28.6)
Histological grade			0.000			0.920			0.894
Well differentiated	4 (10.3)	27 (34.2)		2 (11.8)	11 (31.4)		4 (16.0)	13 (26.5)
Moderately differentiated	26 (66.7)	49 (62.0)		12 (70.6)	23 (65.7)		18 (72.0)	34 (69.4)
Poorly differentiated	9 (23.1)	3 (3.8)		3 (17.6)	1 (2.9)		3 (12.0)	2 (4.1)

P value represents analysis of PNI-positive and PNI-negative datasets between each group. PNI: Perineural infiltration; CEA: Carcinoembryonic antigen; CA19-9: Carbohydrate antigen 19-9; mCRM: MRI-based circumferential resection margin; mEMVI: MRI-based extramural vascular invasion; mT: MRI T stage; mN: MRI N stage; cM: Clinical M stage; cTNM: Clinical TNM stage; DIS: The distance between tumor and anal margin, Low (0-5 cm from the anal verge), middle (5.1-10 cm from the anal verge), and high (10.1-15 cm from the anal verge).

Table 3 Performance of various predictive models in the training, internal validation and external validation groups

Models	Training group	Internal validation group	External validation group
T2WI
AUC (95%CI)	0.817 (0.733-0.901)	0.763 (0.626-0.900)	0.759 (0.644-0.875)
Sensitivity	0.564	0.294	0.480
Specificity	0.899	0.886	0.857
Positive predictive value	0.733	0.556	0.632
Negative predictive value	0.807	0.721	0.764
T1CE
AUC (95%CI)	0.798 (0.707-0.890)	0.689 (0.521-0.857)	0.841 (0.752-0.930)
Sensitivity	0.487	0.471	0.480
Specificity	0.937	0.857	0.878
Positive predictive value	0.792	0.615	0.667
Negative predictive value	0.787	0.769	0.768
T2WI + T1CE
AUC (95%CI)	0.839 (0.757-0.921)	0.787 (0.650-0.923)	0.836 (0.735-0.937)
Sensitivity	0.641	0.529	0.560
Specificity	0.899	0.914	0.939
Positive predictive value	0.758	0.750	0.824
Negative predictive value	0.835	0.800	0.807
Clinical model
AUC (95%CI)	0.804 (0.727-0.881)	0.828 (0.719-0.937)	0.813 (0.724-0.903)
Sensitivity	0.718	0.706	0.800
Specificity	0.747	0.829	0.694
Positive predictive value	0.583	0.667	0.571
Negative predictive value	0.843	0.829	0.872
CR model
AUC (95%CI)	0.889 (0.824-0.954)	0.889 (0.803-0.976)	0.894 (0.814-0.974)
Sensitivity	0.692	0.647	0.760
Specificity	0.924	0.886	0.899
Positive predictive value	0.818	0.733	0.792
Negative predictive value	0.859	0.838	0.880

T2WI: T2-weighted imaging; T1CE: Contrast-enhanced T1WI; CR: Clinical-radiomics prediction model; AUC: Area under receiver operating characteristic curve.

Table 4 Results of univariate and multivariate logistic regression analyses

Variables	Univariate logistic regression		Multivariate logistic regression
	OR (95%CI)	P value	OR (95%CI)	P value
Gender	0.927 (0.418-2.098)	0.854	NA	NA
Age	0.976 (0.944-1.008)	0.138	NA	NA
CEA	1.802 (0.801-4.045)	0.152	NA	NA
CA19-9	3.086 (1.056-9.375)	0.040a	NA	NA
DIS	1.487 (0.968-2.329)	0.075	NA	NA
mT	1.717 (0.919-3.312)	0.096	NA	NA
mN	2.507 (1.467-4.495)	0.001a	NA	NA
cM	5.662 (1.158-40.921)	0.044a	NA	NA
cTNM	3.705 (2.139-7.056)	0.000a	42.002 (2.913-605.511)	0.006a
mCRM	3.439 (1.562-7.838)	0.003a	NA	NA
mEMVI	2.613 (1.199-5.852)	0.017a	NA	NA
Histological grade	0.229 (0.092-0.496)	0.001a	0.113 (0.020-0.658)	0.015a

^aP < 0.05.

CEA: Carcinoembryonic antigen; CA19-9: Carbohydrate antigen 19-9; DIS: The distance between tumor and anal margin; mT: MRI T stage; mN: MRI N stage; cM: Clinical M stage; cTNM: Clinical TNM stage; mCRM: MRI-based circumferential resection margin; mEMVI: MRI-based extramural vascular invasion; OR: Odds ratio; CI: Confidence interval; NA: Not available.

Table 5 Performance of various radiomics predictive models in the training group as evaluated using the Delong test, integrated discrimination improvement index and net reclassification improvement index

Radiomics prediction models	AUC (95%CI)	Delong test P value	IDI (95%CI)	IDI index P value	NRI (95%CI)	NRI index P value
T2WI + T1CE	0.839 (0.757-0.921)
T2WI	0.817 (0.733-0.901)	0.252	0.081 (0.031-0.131)	0.001	0.510 (0.104-0.865)	0.008
T1CE	0.798 (0.707-0.890)	0.196	0.127 (0.064-0.190)	0.000	0.536 (0.255-1.018)	0.005

AUC: Area under receiver operating characteristic curve; IDI: Integrated discrimination improvement; NRI: Net reclassification improvement; T2WI: T2-weighted imaging; T1CE: Contrast-enhanced T1WI.

Table 6 Performance of the clinical model, T2-weighted imaging + contrast-enhanced T1WI fusion sequence radiomics model, and clinical-radiomics model in the training group, as evaluated using the Delong test, integrated discrimination improvement index, and Net reclassification improvement index

Models	AUC (95%CI)	Delong test P value	IDI (95%CI)	IDI index P value	NRI (95%CI)	NRI index P value
CR model	0.889 (0.824-0.954)
Clinical	0.804 (0.727-0.881)	0.009	0.210 (0.130-0.290)	0.000	0.588 (0.271-0.904)	0.000
T2WI + T1CE	0.839 (0.757-0.921)	0.019	0.075 (0.026-0.124)	0.002	0.447 (0.164-0.731)	0.002

AUC: Area under receiver operating characteristic curve; IDI: Integrated discrimination improvement; NRI: Net reclassification improvement; CR: Clinical-radiomics; T2WI: T2-weighted imaging; T1CE: Contrast-enhanced T1WI.