Deep learning model combined with computed tomography features to preoperatively predicting the risk stratification of gastrointestinal stromal tumors

Table 1 Distribution of the analyzed clinical-imaging features across pathologic risk categories, mean ± SD

CECT features		Pathologic risk categories			P value
		High (n = 213)	Moderated (n = 143)	Low (n = 195)
Gender	Male	107	60	89	0.296
	Female	106	83	106
Location	Gastric	139	136	171	0.000
	Non-gastric	74	7	24
Morphology	Regular	54	94	170	0.000
	Irregular	159	49	25
Growth pattern	Endoluminal	35	54	93	0.000
	Mixed	51	22	28
	Exophytic	127	67	74
Degree of contrast enhancement in the venous phase	Mild (≤ 20 HU)	30	20	20	0.424
	Moderate (20-40 HU)	101	74	90
	Obvious (≥ 40 HU)	82	49	85
Contrast enhancement pattern during the venous phase	Continuous	198	142	193	0.000
	Attenuation	15	1	2
Calcification	Present	22	16	21	0.967
	Absent	191	127	174
Necrosis	Present	155	65	33	0.000
	Absent	58	78	162
Ulceration	Present	53	23	8	0.000
	Absent	160	120	187
Enlarged feeding vessels	Present	183	60	14	0.000
	Absent	30	83	181
Lymph nodes	Present	19	1	1	0.000
	Absent	194	142	194
Age		59.44 ± 10.41	61.39 ± 9.83	60.37 ± 9.13	0.183
Size		9.03 ± 4.42	4.90 ± 1.89	2.77 ± 1.20	0.000
Range of tumor enhancement during the arterial phase		18.83 ± 17.81	14.24 ± 11.53	18.45 ± 17.47	0.004
Range of tumor enhancement during the venous phase		41.59 ± 25.73	38.97 ± 18.96	45.22 ± 25.70	0.038
Range of tumor enhancement during the delay phase		42.73 ± 18.89	43.72 ± 17.97	47.97 ± 20.27	0.016

CECT: Contrast-enhanced computed tomography.

Table 2 Logistic regression analysis of risk classification based on clinical-imaging feature

	P value	95%CI
		Lower bound	Upper bound
Size	0.000	-0.763	-0.473
Range of tumor enhancement during the arterial phase	0.131	-0.035	0.005
Range of tumor enhancement during the venous phase	0.220	-0.007	0.032
Range of tumor enhancement during the delay phase	0.858	-0.021	0.017
Morphology	0.602	-0.387	0.608
Location	0.074	-0.063	1.386
Ulceration	0.004	-1.622	-0.300
Enlarged feeding vessels	0.000	-2.134	-1.094
Growth pattern	0.224	-0.833	0.328
Contrast enhancement during the venous phase	0.428	-2.266	1.384
Necrosis	0.236	-0.195	0,793
Lymph nodes	0.890	-1.934	1.678

Table 3 Different algorithms for predicting gastrointestinal stromal tumor risk classification

Different method	Accuracy (%)	Sensitivity (%)	Specificity (%)	F1 score (%)
3DCNN	60	51	68	52
3DResnet_50	66	58	71	59
3DResnet_18	71	66	76	67
3DResnet_34	75	72	87	72
Combined model (3DResnet + MLP)	84	83	92	83

3DCNN: Three-dimensional convolutional neural network; 3DResnet: Three-dimensional residual network; MLP: Multilayer perceptron.

Table 4 Accuracy, sensitivity, specificity, F1 score and areas under the receiver operating characteristic curves for each tumor risk stratification, n (%)/95%CI

		Accuracy (n = 165)	Sensitivity	Specificity	F1 score (%)	AUROC
DLM	High	138 (84); (78-90)	81 (52/64); (76-86)	85 (86/101); (78-91)	79	0.90 (85-95)
	Moderate	131 (79); (71-87)	50 (21/42); (25-74)	89 (110/123); (82-96)	55	0.74 (67-81)
	Low	144 (87); (81-93)	86 (51/59); (81-91)	88 (93/106); (83-94)	83	0.96 (94-98)
	Overall	75	72	87	72	0.88 (83-93)
Combined model	High	148 (90); (86-94)	88 (56/64); (83-93)	91 (92/101); (83-98)	87	0.96 (94-98)
	Moderate	143 (87); (83-92)	69 (29/42); (66-71)	93 (114/123); (86-98)	72	0.83 (78-88)
	Low	152 (92); (85-96)	92 (54/59); (89-93)	92 (98/106); (86-97)	89	0.97 (96-98)
	Overall	84	83	92	83	0.94 (93-95)

DLM: Deep learning model; AUROC: Areas under the receiver operating characteristic curve.