A predictive system comprising serum microRNAs and radiomics for residual retroperitoneal masses in metastatic nonseminomatous germ cell tumors

Summary

Predicting the histopathology of residual retroperitoneal masses (RMMs) before post-chemotherapy retroperitoneal lymph node dissection in metastatic nonseminomatous germ cell tumors (NSGCTs) can guide individualized treatment and minimize complications. Previous single approach-based models perform poorly in validation. Herein, we introduce a machine learning model that evolves from a single-dimensional tumor diameter to incorporate high-dimensional radiomic features, with its effectiveness assessed using the macro-average area under the receiver operating characteristic curves (AUCs). In addition, we utilize more precise and specific microRNAs (miRNAs), not common clinical indicators, to construct an integrated radiomics-miRNA predictive system, achieving an AUC of 0.91 (0.80–0.99) in the prospective test set. We further develop a web-based dynamic nomogram for swift and precise calculation of the histopathological probabilities of RMMs based on radiomic scores and serum miRNA levels. The radiomics-miRNA integrated system offers a promising tool to select personalized treatments for patients with metastatic NSGCT.

Graphical abstract

Highlights

• •A radiomics-miRNA model enables prediction of the histopathology of pcRPLND
• •The integrated model shows better predictive accuracy than the single methodology
• •A web-based tool enables swift calculation of histopathological probabilities

Li et al. develop a radiomics-microRNA integrated predictive system, which can predict the histopathology of residual retroperitoneal masses before post-chemotherapy retroperitoneal lymph node dissection in metastatic nonseminomatous germ cell tumors. The utilization of microRNAs significantly enhances the predictive performance of the integrated model, demonstrating its potential for personalized treatment.

Introduction

Patients with metastatic nonseminomatous germ cell tumor (NSGCT) often have with residual retroperitoneal masses (RRMs) after cisplatin-based chemotherapy.¹ Current guidelines recommend post-chemotherapy retroperitoneal lymph node dissection (pcRPLND) for patients with normal or plateaued serum tumor markers and residual masses >1 cm in the greatest diameter on cross-sectional computed tomography (CT) imaging.² These masses harbor teratomas in around 40% of cases and visible viable germ cell tumors (GCTs) in approximately 10%–15%.³ However, up to 50% of these interventions uncover only benign necrosis or fibrosis, suggesting undergoing possible overtreatment and associated morbidity.⁴ Furthermore, teratomas are resistant to chemotherapy, and cisplatin-based regimens can cause severe side effects; therefore, it is essential to minimize chemotherapy exposure for patients with pure teratomas before surgery.⁵ An accurate prediction model of necrotic/fibrotic tissue and teratoma components can prevent overtreatment, optimize the timing of surgery, and reduce the use of ineffective chemotherapy.

Previous predictive models have often been confined to singular methodologies, such as clinical factors or RRM diameter on imaging. Recent advancements in machine learning and the burgeoning field of radiomics offer potential solutions.^{6,7,8,9,10,11,12} Radiomics involves the conversion of medical images into mineable high-dimensional data, and subsequent quantitative analysis based on machine learning can support decision-making.^10,13 Preliminary studies used radiomics to differentiate benign (necrosis/fibrosis) from malignant (teratoma or viable GCT) tissue, achieving an accuracy rate of up to 81%.⁶ However, challenges persist, especially in discerning teratomas from viable GCTs in post-chemotherapy masses. This limitation, combined with poor interpretability, which affects the ability of radiomics to function as a standalone predictive technique, underscores the necessity for integrated diagnostic approaches.^14,15

MicroRNAs (miRNAs) are small noncoding RNAs involved in cell proliferation and survival.^{16,17,18,19,20} MiR-371a-3p (miR371) has shown promise to accurately identify small-volume visible GCTs from benign processes or pure teratomas before retroperitoneal lymph node dissection (RPLND).^21,22,23 However, miR371 is not expressed in teratomas; therefore, relying solely on this marker might not suffice for treatment decision-making.^24,25 MiR-375-3p (miR375) has been previously studied for its potential in differentiating teratomas; however, its effectiveness remains controversial.²⁵ Studies indicate that integrating circulating miR371 with miR375 could enhance the prediction of GCT components, but there is a lack of prospective cohort validation.^26,27 Additionally, our previous discussion suggests that leveraging these molecular markers alongside radiomics has the potential to revolutionize treatment decision-making and prevent overtreatment.²⁸

This study hypothesized that a predictive system incorporating serum miRNAs and radiomics could more effectively identify patients with NSGCT suitable for pcRPLND and those who might benefit from alternative treatments compared with single-factor prediction methods. The primary objective was to develop a radiomics-miRNA integrated machine learning model utilizing a robust database of CT imaging and serum specimens to accurately determine the histopathological nature of RRMs in patients with NSGCT, and to further validate the model through a prospective test set.

Results

Patient characteristics

A total of 167 patients who underwent RPLND at single institution from September 2014 to August 2022 were retrospectively considered, among which 142 patients underwent cisplatin-based chemotherapy, followed by pcRPLND for RRMs >1 cm. After excluding patients because of incomplete data (missing CT images or serum samples) and 9 for various other reasons (e.g., localization issues, poor image quality), we finalized a retrospective set of 105 male patients with 145 RRMs. The dataset was randomly divided into separate training and validation sets. Additionally, we conducted prospective research involving 45 patients who underwent RPLND treatment at our center between October 2022 and June 2024. After applying the same inclusion and exclusion process as the retrospective study, a total of 34 patients with 42 RRMs were included in the prospective cohort. Figure 1 details the patient enrollment process for this study.

Figure 1.

Flowchart of patient inclusion and exclusion

GCT, germ cell tumor; pcRPLND, post-chemotherapy retroperitoneal lymph node dissection.

These retrospective patients represented all three pathological outcomes from pcRPLND: necrosis/fibrosis (47 patients [44.8%]; 68 masses), teratoma (34 patients [32.4%]; 45 masses), and viable GCT (24 patients [22.8%]; 32 masses). The median age was 26 (interquartile range [IQR], 23–32) years. The prospective cohort showed 14 patients (41%) with necrosis/fibrosis, 15 patients (44%) with teratoma, and 5 patients (15%) with viable GCT (median age, 30 years [IQR, 25–34 years]). To reduce data imbalance, 14, 15, and 13 masses were included for each group, respectively. The characteristics of patients are listed in Tables 1 and 2.

Table 1.

Patient characteristics and clinicopathological data of the retrospective cohort

Characteristic	Total (n = 105)	Necrosis/fibrosis (n = 47)	Teratoma (n = 34)	Viable GCT (n = 24)	p Valueb
Age at pcRPLND (y)a	26 (23–32)	29 (25–36)	26 (21–29)	24 (22–28)	0.01e
Residual masses	145	68 (47)	45 (31)	32 (22)	–
Clinical stage initial	–	–	–	–	0.25c
I (with relapse)	9 (9)	4 (9)	4 (12)	1 (4)	–
II	32 (30)	10 (21)	11 (32)	11 (46)	–
III	64 (61)	33 (70)	19 (56)	12 (50)	–
IGCCCG initial	–	–	–	–	0.71c
Good	51 (49)	24 (51)	15 (44)	12 (50)	–
Intermediate	32 (30)	16 (34)	10 (29)	6 (25)	–
Poor	22 (21)	7 (15)	9 (27)	6 (25)	–
Serum marker before chemotherapy
AFP (ng/mL)a	80.1 (3.4–589.3)	7.6 (3.3–335.4)	263.2 (95.5–1683.2)	11.4 (2.2–244.7)	0.005e
βHCG (mIU/L)a	142.1 (3.0–1261.0)	131.3 (3.6–507.5)	471.4 (29.9–2718.0)	10.3 (0.2–1013.7)	0.09
LDH (U/L)a	339.8 (194.5–854.8)	350.9 (241.4–916.8)	295.5 (160.4–519.4)	410.4 (228.4–883.6)	0.27
Serum marker before pcRPLND
AFP (ng/mL)a	3.5 (2.7–5.8)	3.9 (2.9–5.6)	3.1 (2.02–5.44)	3.3 (2.8–9.1)	0.28
βHCG (mIU/L)a	0.2 (0.1–0.4)	0.2 (0.1–0.2)	0.2 (0.10–0.21)	0.2 (0.2–3.6)	0.10
LDH (U/L)a	171.9 (144.7–193.4)	171.9 (149.8–191.7)	174.2 (146.45–206.58)	162.5 (143.9–188.0)	0.85
Primary histopathology	–	–	–	–	0.07c
Gonadal	96 (91)	42 (89)	34 (100)	20 (83)	–
Extragonadal	9 (9)	5 (11)	0 (0)	4 (17)	–
Tumor elements present in primary tumors
Seminoma	47 (45)	28 (60)	11 (32)	8 (33)	0.02c,e
Embryonal carcinoma	60 (57)	23 (49)	24 (71)	13 (54)	0.14c
Yolk sack tumor	37 (35)	10 (21)	17 (50)	10 (42)	0.02c,e
Chorion carcinoma	9 (9)	2 (4)	4 (12)	3 (12)	0.36c
Teratoma	37 (35)	9 (19)	20 (59)	8 (33)	0.001c,e
Chemotherapy regimen	–	–	–	–	<0.001c,e
First-line	75 (71)	41 (87)	25 (74)	9 (38)	–
First-line + salvage	30 (29)	6 (13)	9 (26)	15 (62)	–
Type of first-line chemotherapy	–	–	–	–	0.03c,e
BEP	79 (75)	35 (74)	29 (85)	15 (62)	–
EP	10 (9)	6 (13)	0 (0)	4 (17)	–
BEP + EP	3 (3)	3 (6)	0 (0)	0 (0)	–
VIP	8 (8)	2 (4)	5 (15)	1 (4)	–
Others	5 (5)	1 (2)	0 (0)	4 (17)	–
First-line chemotherapy cycles	–	–	–	–	<0.001c,e
2	10 (9)	5 (11)	4 (12)	1 (4)	–
3	23 (22)	12 (26)	7 (21)	4 (17)	–
4	62 (59)	27 (57)	20 (59)	15 (62)	–
5	10 (9)	3 (6)	3 (9)	4 (17)	–
Type of salvage chemotherapy	–	–	–	–	0.69c
TIP	16 (53)	3 (50)	6 (67)	7 (47)	–
VIP	3 (10)	0 (0)	1 (11)	2 (13)	–
VeIP	1 (3)	0 (0)	0 (0)	1 (7)	–
VIP + TIP	3 (10)	2 (33)	0 (0)	1 (7)	–
VAB-6	1 (3)	0 (0)	1 (11)	0 (0)	–
Others	6 (20)	1 (17)	1 (11)	4 (27)	–
Time from last day of chemotherapy to pcRPLND (d)a	54 (43–65)	54 (38–64)	58 (49–72)	52 (36–64)	0.58
Residual mass size on CT before chemotherapy (mm)a	49 (26–73)	52 (27–70)	46 (34–77)	28 (25–65)	0.21
Residual mass size on CT before pcRPLND (mm)a	29 (21–46)	26 (20–39)	44 (30–72)	26 (21–34)	0.001e
Change in mass size (mm)a,d	10 (−2 to 32)	22 (5–36)	2 (−5 to 24)	0 (−2 to 26)	0.001e

Except where indicated, data are numbers of patients, with percentages in parentheses.

GCT, germ cell tumor; pcRPLND, post-chemotherapy retroperitoneal lymph node dissection; IGCCCG, The International Germ Cell Consortium Classification Prognostic Groups; AFP, alpha-fetoprotein; βHCG, beta human chorionic gonadotropin; LDH, lactate dehydrogenase; BEP, bleomycin, etoposide, and cisplatin; EP, etoposide and cisplatin; VIP, cisplatin, etoposide, and ifosfamide; TIP, paclitaxel, ifosfamide, and cisplatin; VeIP, vinblastine, ifosfamide, and cisplatin; VAB-6, cyclophosphamide, vinblastine, bleomycin, dactinomycin, and cisplatin.

^aData are medians, with interquartile ranges (IQRs) in parentheses.

^bp values from Kruskal–Wallis H-test, except where otherwise indicated.

^cp value from Pearson chi-squared test or Fisher’s exact test.

^dBefore chemotherapy minus after chemotherapy.

^eIndicates statistical significance; p < 0.05.

Table 2.

Patient characteristics and clinicopathological data of the prospective cohort

Characteristic	Total (n = 34)	Necrosis/fibrosis (n = 14)	Teratoma (n = 15)	Viable GCT (n = 5)	p Valueb
Age at pcRPLND (y)a	30 (25–34)	35 (30–37)	27 (22–30)	29 (28–31)	0.001e
Residual masses	42	14 (33)	15 (36)	13 (31)	–
Clinical stage initial	–	–	–	–	0.03c,e
I (with relapse)	0 (0)	0 (0)	0 (0)	0 (0)	–
II	16 (47)	6 (43)	10 (67)	0 (0)	–
III	18 (53)	8 (57)	5 (33)	5 (100)	–
IGCCCG initial	–	–	–	–	0.08c
Good	19 (56)	8 (57)	11 (73)	0 (0)	–
Intermediate	12 (35)	5 (36)	3 (20)	4 (80)	–
Poor	3 (9)	1 (7)	1 (7)	1 (20)	–
Serum marker before chemotherapy
AFP (ng/mL)a	133.1 (5.6–430.2)	7.2 (5.2–159.8)	259.0 (84.2–424.5)	1869.0 (88.9–2463.0)	0.25
βHCG (mIU/L)a	23.2 (0.4–934.8)	1.9 (0.2–23.45)	68.2 (4.0–834.5)	13780.0 (1134.0–25878.0)	0.02e
LDH (U/L)a	230.4 (182.2–389.3)	298.2 (189.1–395.2)	219.5 (182.9–271.7)	388.4 (168.3–554.0)	0.66
Serum marker before pcRPLND
AFP (ng/mL)a	4.7 (3.1–6.3)	5.0 (3.9–8.0)	4.3 (2.3–5.8)	5.0 (4.0–6.1)	0.39
βHCG (mIU/L)a	0.2 (0.2–0.2)	0.2 (0.2–0.2)	0.2 (0.20–0.34)	0.2 (0.2–3.14)	0.12
LDH (U/L)a	180.9 (157.6–208.9)	173.0 (160.1–206.4)	185.8 (156.4–192.6)	242.7 (180.3–253.7)	0.44
Primary histopathology	–	–	–	–	0.25c
Gonadal	32 (94)	13 (93)	15 (100)	4 (80)	–
Extragonadal	2 (6)	1 (7)	0 (0)	1 (20)	–
Tumor elements present in primary tumors
Seminoma	8 (24)	5 (36)	2 (13)	1 (20)	0.36c
Embryonal carcinoma	19 (56)	5 (36)	10 (67)	4 (80)	0.12c
Yolk sack tumor	14 (41)	4 (29)	6 (40)	4 (80)	0.13c
Chorion carcinoma	3 (9)	1 (7)	2 (13)	0 (0)	0.63c
Teratoma	16 (47)	4 (29)	10 (67)	2 (40)	0.11c
Chemotherapy regimen	–	–	–	–	0.002c,e
First-line	27 (79)	13 (93)	13 (87)	1 (20)	–
First-line + salvage	7 (21)	1 (7)	2 (13)	4 (80)	–
Type of first-line chemotherapy	–	–	–	–	0.69c
BEP	24 (70)	10 (71)	11 (73)	3 (60)	–
EP	3 (9)	1 (7)	2 (13)	0 (0)	–
BEP + EP	4 (12)	2 (14)	1 (7)	1 (20)	–
VIP	2 (6)	0 (0)	1 (7)	1 (20)	–
Others	1 (3)	1 (7)	0 (0)	0 (0)	–
First-line chemotherapy cycles	–	–	–	–	0.89c
2	3 (9)	1 (7)	2 (13)	0 (0)	–
3	11 (32)	6 (43)	4 (27)	1 (20)	–
4	15 (44)	5 (36)	7 (47)	3 (60)	–
5	5 (15)	2 (14)	2 (13)	1 (20)	–
Type of salvage chemotherapy	–	–	–	–	0.18c
TIP	2 (29)	0 (0)	0 (0)	2 (50)	–
VIP	1 (14)	1 (100)	0 (0)	0 (0)	–
VeIP	1 (14)	0 (0)	1 (50)	0 (0)	–
VIP + TIP	2 (29)	0 (0)	1 (50)	1 (25)	–
VAB-6	1 (14)	0 (0)	0 (0)	1 (25)	–
Others	0 (0)	0 (0)	0 (0)	0 (0)	–
Time from last day of chemotherapy to pcRPLND (d)a	51 (40–62)	49 (33–59)	48 (38–57)	54 (38–66)	0.78
Residual mass size on CT before chemotherapy (mm)a	44 (30–64)	58 (37–67)	34 (30–53)	17 (13–54)	0.13
Residual mass size on CT before pcRPLND (mm)a	26 (18–36)	24 (13–28)	33 (24–52)	29 (20–33)	0.13
Change in mass size (mm)a,d	10 (3–24)	28 (21–38)	3 (2–7)	−6 (−7 to 10)	<0.001e

Except where indicated, data are numbers of patients, with percentages in parentheses.

GCT, germ cell tumor; pcRPLND, post-chemotherapy retroperitoneal lymph node dissection; IGCCCG, The International Germ Cell Consortium Classification Prognostic Groups; AFP, alpha-fetoprotein; βHCG, beta human chorionic gonadotropin; LDH, lactate dehydrogenase; BEP, bleomycin, etoposide, and cisplatin; EP, etoposide and cisplatin; VIP, cisplatin, etoposide, and ifosfamide; TIP, paclitaxel, ifosfamide, and cisplatin; VeIP, vinblastine, ifosfamide, and cisplatin; VAB-6, cyclophosphamide, vinblastine, bleomycin, dactinomycin, and cisplatin.

^aData are medians, with interquartile ranges (IQRs) in parentheses.

^bp values from Kruskal–Wallis H-test, except where otherwise indicated.

^cp value from Pearson chi-squared test or Fisher’s exact test.

^dBefore chemotherapy minus after chemotherapy.

^eIndicates statistical significance; p < 0.05.

According to univariate analysis, in the retrospective cohort, clinicopathological characteristics, including age; serum alpha-fetoprotein measured before chemotherapy; the presence of seminoma, yolk sack tumor, and teratoma in the primary tumors; chemotherapy regimen; preoperative mass size; and post-chemotherapy change in mass size, were significantly associated with the histopathology of pcRPLND. In the prospective cohort, significant associations with the pathological outcomes of pcRPLND were found for age, clinical stage at diagnosis, pre-chemotherapy beta human chorionic gonadotropin levels, chemotherapy regimen, and post-chemotherapy change in mass size.

The dataset was randomly divided into a training set (75 patients; 108 masses, n = 51 necrosis/fibrosis, n = 33 teratoma, and n = 24 viable GCT; median age, 26 years [IQR, 22–32 years]) and a validation set (30 patients; 37 masses, n = 17 necrosis/fibrosis, n = 12 teratoma, and n = 8 viable GCT; median age, 26 years [IQR, 23–30 years]), ensuring no significant differences in baseline demographics (Table S1). There were also no significant differences in general and clinicopathological data between the retrospective and prospective cohorts (Table S2).

Dimension reduction and radiomics feature selection

A total of 1,130 features were collected from the region of interest (ROI) of each residual mass. After assessing reproducibility using intra- and interobserver agreement, 850 features were retained. Further statistical analysis in the training set identified 223 features that were significantly associated with the histopathological classes. Highly correlated features (Pearson’s r ≥ 0.8) were identified using correlation matrices (Figure S1A), resulting in the exclusion of 190 features. The application of least absolute shrinkage and selection operator (LASSO) regression to the remaining 33 features identified six optimal features for the radiomic model (Figures S2A and S2B). Additionally, weight and heat maps for these six features were produced (Figures S2C and S2E). The boxplots for the six features across the training set and test set are shown in Figures S2D and S2F. The Radscore for each residual mass was calculated using the following formula:

$$\text{Radscore}=0.815+0.102\ast \text{original}\_\text{shape}\_\text{Flatness}+0.024\ast \text{original}\_\text{ngtdm}\_\text{Contrast}-0.018\ast \text{wavelet}\_\text{LHL}\_\text{glcm}\_\text{Imc}2-0.087\ast \text{wavelet}\_\text{LHL}\_\text{gldm}\_\text{DependenceVariance}+0.094\ast \text{wavelet}\_\text{LLL}\_\text{gldm}\_\text{DependenceEntropy}-0.034\ast \log \_\text{sigma}\_3\_\text{mm}\_3\mathrm{D}\_\text{glszm}\_\text{SmallAreaEmphasis.}$$

Similarly, the same analytical approach yielded ten features that could distinguish “benign” (necrosis/fibrosis) from “malignant” (teratoma or viable GCT) cases. Details on feature filtering are presented in Figure S1B and Figure S3.

Machine learning radiomics models to predict the histopathology of RRMs

The performance metrics of six radiomic models are summarized in Table S3, with their receiver operating characteristic (ROC) curves presented in Figures 2A and 2B. Among them, the logistic regression model showed superior performance. In the validation set, the logistic regression model attained a macro-average area under the receiver operating characteristic curve (AUC) of 0.76 (95% confidence interval (CI): 0.59–0.94) and an accuracy of 0.65 (Table S3; Figures 2C and 2D). Specifically, the predictive abilities for different histopathological subtypes were as follows: AUC for necrosis/fibrosis: 0.78 [0.63–0.94]; AUC for teratoma: 0.86 [0.72–1.00]; and AUC for viable GCT: 0.64 [0.42–0.88]). The confusion matrix of the radiomic model is depicted in Figures 2E and 2F. Besides, when applying the logistic regression model to distinguish between “benign” and “malignant,” it achieved an AUC of 0.90 (95% CI: 0.79–1.00) and an accuracy of 0.81 in the validation set, aligning with findings from previous studies (Table S4).⁶

Figure 2.

Evaluation of the radiomics models

(A and B) The macro-average receiver operating characteristic curves for six supervised radiomics machine learning models in the training and validation sets.

(C and D) The receiver operating characteristic curves for regularized logistic regression model with best performance in the training and validation sets.

(E and F) The confusion matrix for regularized logistic regression model with best performance in the training and validation sets.

N, necrosis/fibrosis; T, teratoma; V, viable germ cell tumor; O, other histopathological classes; AUC, area under the receiver operating characteristic curve.

Specific expression of miRNAs in the serum of patients with NSGCT before pcRPLND

In the training, validation, and test sets, serum miR371 levels were significantly elevated in patients with residual viable GCTs compared with those with necrosis/fibrosis or teratomas (Figures 3A–3C). Conversely, miR375 showed higher expression in patients with teratomas, while it was undetectable in patients with necrosis/fibrosis and viable GCTs (Figures 3D–3F). Employing a logistic regression model for the integrated miR371-miR375 analysis, the test dataset exhibited a macro-average AUC of 0.89 (95% CI: 0.770.97; Figures S4A and S4B), surpassing the individual macro-average AUCs of miR371 and miR375 (0.74, 95% CI: 0.57–0.88 and 0.79, 95% CI: 0.63–0.94, respectively; Figures S4C and S4D). The optimal cutoff values for the individual miRNA models to identify viable GCTs and teratomas were determined as 1.76 for miR371 and 1.32 for miR375.

Figure 3.

Expression and predictive performance of miR-371a-3p and miR-375-5p

(A–C) Expression of miR-371a-3p in patients enrolled in the training, validation, and test sets.

(D–F) Expression of miR-375-5p in patients enrolled in the training, validation, and test sets.

Welch’s analysis of variance (ANOVA) and the Kruskal-Wallis H test with Dunnett’s T3 multiple comparison correction were used. GCT, germ cell tumor; ns, not significant. ∗p < 0.05.

Development and validation of the radiomics-miRNAs integrated model

We incorporated serum miR371 and miR375 into the radiomic model to create an integrated model. This model achieved a macro-average AUC of 0.98 (95% CI: 0.96–1.00) in the training set, 0.96 (95% CI: 0.87–1.00) in the validation set, and 0.91 (95% CI: 0.80–0.99) in the test set (Figures 4A–4C, S5A–S5C). The comparative performance of different models is shown in Table S5 and Figures S5G–S5L. Notably, the integrated model demonstrated superior performance, misclassifying only two necrosis/fibrosis, four teratoma, and three viable GCT cases in the test set (Figures S5C and S5D). The Delong test revealed that the AUC of the integrated model was significantly higher than that of the radiomics-only model in all three datasets (p values = 0.01, 0.03, and 0.04 after Bonferroni correction). Calibration curves indicated better predictive agreement in the integrated model, as evidenced by a lower Brier score (0.08) and log loss (0.42) in the validation set (Figures 4D–4F; Table S6). The Hosmer-Lemeshow test also confirmed a good fit (p > 0.05, Table S6). Additionally, the integrated model showed the lowest Akaike information criterion (AIC) value, suggesting that the combination of imaging and serum markers provided a synergistic improvement in prediction performance (Table S7).

Figure 4.

Evaluation of the radiomics-miRNAs integrated model

(A–C) The macro-average receiver operating characteristic curves for individual radiomics and radiomics-miRNAs integrated models in patients in the training, validation, and test sets.

(D) Calibration of the prediction model in terms of the agreement between predicted and observed respondents between the necrosis/fibrosis and others. The diagonal dotted line represents a perfect prediction; the blue and red lines represent the predictive performance of the individual radiomics and radiomics-miRNA integrated models, respsectively. The closer the model line fit is to the ideal line, the better the predictive accuracy of the model.

(E) Calibration of the prediction model in terms of the agreement between predicted and observed respondents between the teratoma and others.

(F) Calibration of the prediction model in terms of the agreement between predicted and observed respondents between the viable germ cell tumor and others.

(G–I) Decision curves for predicting necrosis/fibrosis, teratomas, and viable germ cell tumors. The black line represents the assumption that all patients have a teratoma or viable germ cell tumor, and the horizontal dotted line represents the assumption that no residual teratomas or viable germ cell tumors remained in all patients. The blue area represents the benefits of an individual radiomics model in clinical decision-making. The red area represents the greater benefits of the integrated model in clinical use compared with the radiomics model.

T, teratoma; V, viable germ cell tumor; O, other histopathological classes; AUC, area under the receiver operating characteristic curve.

Clinical use of the radiomics-miRNAs integrated predictive system

Decision curve analysis was employed to compare the benefits of the radiomic and integrated models in the clinical decision-making for surgery and chemotherapy. We found that the integrated model consistently showed a superior performance in terms of patient benefit (Figures 4G–4I). Multivariate analysis of Radscores and miRNA levels revealed related significant predictors (Table S8), and three visualized nomograms were constructed to predict each class of histopathology (Figures S6A–S6C). To increase the transparency and practicality of the nomograms to facilitate information sharing, a web application was developed (using the R module Shiny) to interactively explore all the models’ details, which successfully achieved rapid and accurate calculation of the probabilities of the histopathology of RRMs using the radiomic score and serum miRNA levels (Figure S6D, Shinyapp: https://dingrj.shinyapps.io/pcRPLND_prediction). Figure 5 shows data for three patients with similar clinical characteristics, whose post-chemotherapy residual masses manifested distinct histopathological types. Our model accurately predicted the histopathology of RRMs in these patients, potentially reducing the risk of overtreatment.

Figure 5.

Schematic showing example implementation of the integrated radiomics-miRNA model in three patients with a metastatic NSGCT with RRMs after cisplatin-based chemotherapy who have similar clinical characteristics

Patient 15, patient 27, and patient 102, diagnosed at the approximate ages of 22, 26, and 23 years, respectively, had the same clinical stages of T2N1M0S1IIa and were classified as having good prognosis according to The International Germ Cell Consortium Classification Prognostic Groups (IGCCCG) before chemotherapy. Patient 15 was diagnosed with left testicular mixed germ cell tumor (yolk sac tumor 40%, teratoma 40%, embryonal carcinoma 15%, and seminoma 5%). Before chemotherapy, the tumor markers alpha-fetoprotein (AFP), beta human chorionic gonadotropin (βHCG), and lactate dehydrogenase (LDH) were measured at 97.4 ng/mL, 45.6 mIU/L, and 137.6 U/L, respectively. Following four cycles of bleomycin, etoposide, and cisplatin (BEP) chemotherapy, the patient’s tumor markers returned to normal, and the diameter of the RRM decreased from 25 mm before chemotherapy to 17 mm after chemotherapy. Patient 27 was diagnosed with right testicular mixed NSGCT (embryonal carcinoma 75%, teratoma 20%, choriocarcinoma 4%, and yolk sac tumor 1%) before chemotherapy, with AFP, βHCG, and LDH at 11.7 ng/mL, 399 mIU/L, and 151.7 U/L, respectively. After three cycles of BEP chemotherapy, the patient’s tumor markers normalized, and the tumor shrank from 27 mm to 21 mm. Patient 102 was diagnosed with right testicular mixed NSGCT (yolk sac tumor 60%, embryonal carcinoma 40%) who exhibited pre-chemotherapy tumor marker levels of 413 ng/mL, 634 mIU/L, and 378.1 U/L for AFP, βHCG, and LDH, respectively. After three cycles of BEP chemotherapy, despite normalized tumor markers, there was little change in the residual mass (from 25 mm to 24 mm). We manually delineated the RRM on post-chemotherapy contrast-enhanced abdominal computed tomography (CT) scans. Subsequently, we extracted features from the venous phase of the images and computed the Radscores using our radiomics model (68.2 for patient 15, 91.5 for patient 27, and 105.3 for patient 102). Concurrently, we extracted microRNAs from post-chemotherapy peripheral serum, including miR-371a-3p (miR371) and miR-375-5p (miR375) and calculated their relative expression normalized to cel-miR-39-3p and miR-30b-5p (for patient 15: miR371 = 0.03 and miR375 = 0.85; for patient 27: miR371 = 0.01 and miR375 = 8.98; for patient 102: miR371 = 16.75 and miR375 = 0.09). By integrating Radscores and miRNA expression, we identified the most probable histopathological types for the residual masses in the three patients. The pathology after retroperitoneal lymph node dissection confirmed our predictions (magnification, ×10; scale bars, 200 μm). Ultimately, patients 15 and 27 showed no recurrence during follow-up. Patient 102 underwent two cycles of paclitaxel, ifosfamide, and cisplatin (TIP) adjuvant chemotherapy, with no observed recurrence or metastasis at present. P_N, probability of predicting necrosis/fibrosis; P_T, probability of predicting a teratoma; P_V, probability of predicting a viable germ cell tumor.

Discussion

Herein, we developed an integrated radiomics-miRNA machine learning model with the aim of accurately characterizing the histopathology of RRMs in patients with NSGCT before pcRPLND (macro-average AUC in the validation set, 0.96 [0.87–1.00]). Then the histopathology classifier of pcRPLND has been validated prospectively in the test set with a macro-averaged AUC of 0.91 (0.80–0.99). Our multi-classification approach potentially allows up to half of the patients, who only have residual necrosis/fibrosis, to avoid unnecessary pcRPLND. Moreover, since cisplatin-based chemotherapy results in severe toxicity, such as nephrotoxicity and pulmonary toxicity, as well as because of its impact on fertility, our model provides the possibility of identifying patients with residual chemo-refractory pure teratomas, thereby potentially reducing unnecessary chemotherapy exposure before RPLND.⁵

Radiomics enhances the quantitative analysis of medical images and aims to support clinical decision-making by overcoming the constraints of purely visual interpretation.^9,10,29 Radiomics is particularly useful to evaluate RRMs in GCT, considering the morphological heterogeneity exhibited by NSGCTs.⁷ Recent studies have explored CT radiomics-based classifiers to predict the histopathological nature of residual masses before pcRPLND. For instance, Baessler et al., reported a single-center study with 81 patients and 204 ROIs, which achieved an accuracy of 81% to distinguish benign from malignant histopathology, which is the same as our study’s accuracy of 81%.⁶ Unlike these studies, we differentiated among three tissue types and still attained a macro-average AUC of 0.76 (0.59–0.94) using radiomics alone. Most recently, Scavuzzo et al. investigated 122 patients and developed a radiomics-based model with an optimal average AUC of 0.95 using support vector machine (SVM), which is comparable to the AUC of our integrated model (0.96 [0.87–1.00]).⁸ Notably, they trained their model on the entire set without a separate test dataset, potentially raising concerns about overfitting. Although cross-validation was used, the model’s robustness might be affected by the large number of features (30 in total). The use of independent and prospective test sets in the present study enhances confidence in the model’s generalizability and applicability in external test settings.

In our study, we acknowledge that the radiomic signature was heavily influenced by the delineation of ROIs.¹¹ Despite our efforts to evaluate intra- and interobserver intraclass correlation coefficients, the clinical application of a radiomic model could be affected by the variability in CT scanners and observer interpretations. Meanwhile, radiomics might struggle to detect small volumes of viable GCTs within large, predominantly fibrotic or necrotic, lymph node masses because of the limited resolution of CT scans. In our study, differentiating viable GCTs from other subtypes using the radiomic model proved challenging (AUC = 0.64 [0.42–0.88]). Moreover, CT images provide a snapshot of tumor-visible characteristics at a specific time point but do not capture the tumor’s molecular gene expression status, emphasizing the need for molecular markers to enhance prediction accuracy.^17,30,31

Recent research has validated the utility of miR371 to identify patients with viable GCTs.^21,22,23 Reliable biomarkers for mature teratomas would significantly aid clinical decision-making. While the role of miR375 in accurately detecting teratomas remains unclear, our study supports its potential, indicating the need for further research to explore the combined predictive capacity of miR371 and miR375 in larger sets of patients with NSGCT. We observed high expression of miR371 and miR375 in viable GCTs and pure teratomas, respectively. As standalone biomarkers, both miR371 and miR375 showed a modest performance (AUC = 0.74 [0.57–0.88] and 0.79 [0.63–0.94], respectively). When evaluated together, their predictive accuracy for the histopathology of residual masses improved markedly (AUC = 0.89 [0.77–0.97]), aligning with findings from other studies. For example, Nappi et al. reported that an integrated plasma evaluation of miR371-miR375 demonstrated strong performance in detecting teratomas (AUC = 0.95 [0.90–0.99]), albeit with relatively poor performance in the test set (AUC = 0.77 [0.62–0.93]).²⁶ In addition, Kremer et al. reported an excellent AUC of 0.94 (0.87–1.00) to distinguish viable GCTs and teratomas from necrosis/fibrosis by combining miR-371a-3p and miR-375-5p expression levels in postoperative tissue samples.²⁷ However, their study focused on postoperative evaluation rather than predicting histopathology preoperatively and therefore does not enable the avoidance of unnecessary pcRPLND or chemotherapy for patients. Additionally, our study, with the largest patient set (n = 139) to date, further validated the integrated miR371-miR375 model in an prospective dataset (n = 34).

We pioneered the integration of circulating serum miRNA expression data with radiomic features to create a predictive model for metastatic patients with NSGCT. This innovative approach, which combines two distinct methodologies into a statistical model, has shown potential as a robust and noninvasive predictor of histopathology, contributing to the emerging paradigm of personalized treatment.^9,14,32 Our integrated model demonstrated notable predictive ability during testing, achieving a macro-average AUC of 0.91 (0.80–0.99). Moreover, it outperformed the pure radiomics model in terms of predictive accuracy, calibration performance, and clinical utility.³³ To enhance its accessibility and practical application, we developed a web-based nomogram application featuring both the integrated and comparative models (based solely on one of the two signatures), catering for various medical institutions’ capabilities.^34,35 The clinical significance of the integrated model lies in its ability to reduce the effort required by clinicians for comprehensive evaluations by providing more objective and precise guidance. The prediction system is directly applicable to standalone routine enhanced CT images and peripheral blood serum tests, making it easy to implement in primary care settings. In the present study, the multimodal classifier proved to be more accurate and reliable compared with individual modalities, highlighting its superior performance in advanced care settings. Specifically, when clinicians input radiomic scores and miRNA expression data into the integrated model, it provides predicted probabilities for three pathological types. For patients with a high likelihood of a residual pure teratoma, this can help avoid unnecessary chemotherapy or surgery. In these patients, early pcRPLND, rather than continued chemotherapy, can help reduce adverse effects and long-term complications. It is worth mentioning that the expression thresholds for these miRNAs were determined using a laboratory-developed assay to measure their relative levels and are based solely on our predictive model and thus might not be universally applicable. In future research, we will investigate more precise quantitative thresholds and develop clinical testing methods that are better suited for real-world applications.³⁶

Limitations of the study

The present study has several limitations. Firstly, the retrospective nature of the study introduces inherent challenges. The CT images were acquired using diverse scanners and protocols, resulting in a heterogeneous dataset. While this heterogeneity mirrors real-world clinical settings and potentially broadens the model’s applicability, it also presents challenges in ensuring model robustness. Additionally, the relatively high rate (22%) of residual viable disease in this retrospective cohort also reflects selection biases inherent in single-center retrospective studies.^37,38 To mitigate potential overfitting, we employed a rigorous methodology, including an independent prospective test set. The new prospective cohort of 34 patients included 5 patients (15%) with viable GCTs, consistent with the consensus that has already been reported.³⁷ Therefore, under the condition of image registration, we view this data diversity as beneficial for the model’s generalizability rather than a limitation. Secondly, the technical complexity and time-intensive nature of radiomics pose challenges for the implementation of the developed model in clinical settings, particularly outside major tertiary centers. This limitation calls for advances in automated ROI contouring and the application of artificial intelligence.³⁹ Furthermore, for some machine learning classifiers, the substantial performance disparity between the training and validation sets indicates underlying overfitting. To address this, cross-validation and ensemble methods should be employed to assess and mitigate overfitting in future analyses.

In conclusion, the developed radiomics-miRNA integrated predictive system represents a promising tool to improve the prediction of the histopathology of RRMs in cases of metastatic NSGCT, aiming to guide individualized treatment and minimize complications. If confirmed by future evidence, surveillance is adequate for patients predicted to have necrosis/fibrosis, and immediate surgery is recommended for patients with predicted pure teratomas. In contrast, patients with viable GCTs should receive cisplatin-based chemotherapy, followed by re-evaluation to remove masses that still contain malignant components. Supported by these promising results, further external testing and the use of larger prospective datasets are warranted to confirm these findings.

Resource availability

Lead contact

Further information and requests for resources and reagents should be directed to and will be fulfilled by the lead contact, Zhuowei Liu (liuzhw@sysucc.org.cn).

Materials availability

This study did not generate new unique reagents.

Data and code availability

• •The CT image data from this study cannot be publicly deposited due to containing sensitive information that could potentially compromise patient privacy. The dataset in this study is shared on the Research Data Deposit (RDD number are listed in the key resources table). Summary statistics detailing these data have been included in the main and supplementary tables and are publicly accessible as of the publication date.
• •Source code for the machine learning model and data processing is publicly available as of the date of publication. Accession websites are listed in the key resources table.
• •Any additional information required to reanalyze the data reported in this research is available from the lead contact upon request.

Acknowledgments

The funding agencies, including the Cancer Innovative Research Program of Sun Yat-sen University Cancer Center (CIRP-SYSUCC-0014) and Science and Technology Projects in Guangzhou (2023A04J1783), provided foundation but had no role in the research conducted. We are very thankful to all the study participants and acknowledge the core management group for organizing the databases, as well as the participation of experts in reporting of imaging and pathology.

Author contributions

Zhuowei Liu: conceptualization, validation, investigation, resources, data curation, writing – review and editing, supervision, project administration, and funding acquisition. Z.M.: conceptualization, validation, investigation, data curation, writing – review and editing, supervision, project administration, and funding acquisition. X.L.: conceptualization, methodology, formal analysis, investigation, software, data curation, writing – original draft, visualization, project administration, and funding acquisition. R.D.: methodology, software, formal analysis, investigation, data curation, writing – original draft, and visualization. Zhenhua Liu: methodology, software, validation, formal analysis, and data curation. W.M.S.T.: formal analysis and writing – original draft. J.Y.: software, formal analysis, and data curation. L.T.: methodology, data curation, and validation. H.L.: methodology, data curation, and validation. S.G.: validation, writing – review and editing, and supervision. K.Y.: validation, writing – review and editing, and supervision.

Declaration of interests

The authors declare no competing interests.

STAR★Methods

Key resources table

REAGENT or RESOURCE	SOURCE	IDENTIFIER
Biological samples
Antecubital vein serum of NSGCT patients before pcRPLND and age-matched healthy male blood donors	Sun Yat-sen University Cancer Center	Related to Tables 1, 2, S1, and S2
Chemicals, peptides, and recombinant proteins
TRIzol Reagent	Invitrogen	Cat# 15596026
PBS	Life Technologies	Cat# 14190169
Critical commercial assays
Mir-X™ miRNA First Strand Synthesis Kit	TaKaRa	No. 638315
Mir-X™ miRNA qRT-PCR SYBR® Kit	TaKaRa	No. 638313
QuantstudioTM 7 Flex Real-Time PCR	Applied Biosystems	N/A
Oligonucleotides
Primers for qRT-PCR	Ruibiotech	Table S12
Deposited data
Patients dataset	This paper	https://www.researchdata.org.cn; number RDDA2024167073
Software and algorithms
Code	This manuscript	https://github.com/Renjie-D/pcRPLND-prediction.git
SPSS 24.0	IBM	https://www.spss.cm.cn
R 4.1.0	R Project	https://www.r-project.org
Python 3.7.4	Python software	https://www.python.org
GraphPad PRISM 8	GraphPad software	https://www.graphpad.com
Pyradiomics 3.0.1	Griethuysen JJM et al.42	https://pyradiomics.readthedocs.io/en/latest/faq.html
ITK-SNAP 3.8.0	ITK-SNAP software40	http://www.itksnap.org
SimpleITK	SimpleITK41	https://github.com/SimpleITK

Experimental model and subject details

Human participants were included in this study, with the detailed inclusion and exclusion criteria outlined in the Method Details section and illustrated in Figure 1. All the patients enrolled in this study were male, Han Chinese ethnicity, and East Asian ancestry. The demographic characteristics of all patients are presented in Tables 1 and 2. All samples were obtained from Sun Yat-sen University Cancer Center. The sample size justification is shown in the Method Details. The retrospective dataset was randomly divided into separate training and validation sets. All patients from the screened prospective cohort were included as the test set. This study adheres to the Declaration of Helsinki and has received approval from the ethical review authority (B2023-058-01). The requirement for written informed consent from patients was waived in the retrospective cohort. Patients in the prospective cohort were part of an enrolled clinical trial (Chinese Clinical Trail Registry: www.chictr.org.cn, number ChiCTR2400081978), all of whom were recruited at the Sun Yat-sen University Cancer Center and provided written informed consent in accordance with the Declaration of Helsinki. The dataset in this study is shared on the Research Data Deposit (RDD number: RDDA2024167073).

Method details

Study design and patients

The Standards for Reporting of Diagnostic Accuracy Studies (STARD) and CheckList for EvaluAtion of Radiomics research (CLEAR) recommendations were applied for reporting (Methods S1).

The retrospective study sample comprised consecutive patients with metastatic GCTs who underwent RPLND at Sun Yat-sen University Cancer Center from September 2014 to August 2022. The prospective test set involving patients who underwent pcRPLND at Sun Yat-sen University Cancer Center between October 2022 and June 2024. The cohorts used in this study have not been previously reported or published elsewhere. The inclusion criteria were as follows: (a) primary lesion diagnosed as testicular or primary retroperitoneal NSGCT, (b) treated with cisplatin-based chemotherapy, (c) with normal or plateaued serum tumor markers prior to pcRPLND, (d) wih RRM greater than 1cm in greatest diameter on cross-sectional CT imaging after chemotherapy, and (e) underwent standard nerve-sparing pcRPLND. The exclusion criteria were: (a) missing contrast-enhanced abdominal CT image, (b) missing preoperative serum sample, (c) difficulty in localization of retroperitoneal lesion, (d) poor image quality, and (e) difficulty in correlating lesions on imaging with histopathology.

Sample size justification

The determination of an appropriate sample size typically relies on a rule of thumb established from simulation studies, suggesting a minimum of 10 events per variable (EPV).⁴³ Based on previous studies of pcRPLND radiomics cohorts,^6,7,44 which estimated inclusion of 5–8 imaging features in addition to serum miR-371a-3p and miR-375-5p levels, it was anticipated that 7–10 variables would be included in the multifactorial regression model. Following the EPV principle, the minimum sample size required was estimated to be 70–100 cases. This study included 108 retroperitoneal lesions in the training set to establish the model, meeting the sample size requirement.

CT image acquisition

All patients underwent contrast-enhanced CT scans, and informed consent forms were signed before inspection. Contrast-enhanced CT imaging of the abdomen and pelvis was performed with nonionic intravenous contrast. The CT scans were acquired with the six CT machine (see details in Table S9).

Preparation for the examination occurred as follows: patients fasted for more than 8 h before the examination and breath-holding exercises were performed. Thirty minutes prior to the examination, instruct the patient to consume 500-800mL of warm water. During the examination, the patient should lie down facing up with arms raised and feet entering first. The scanning range should extend from the diaphragm to the upper edge of the pubic symphysis. The scanning parameters were as follows: a pitch of 1.000, X-ray tube rotation time of 0.5s/r, collimator set at 40mm, automatic mAs technology employed, tube voltage set to 120kVp, matrix 512 × 512 mm, and a scan thickness of 0.625–5 mm with scan spacing from 0.625 to 5 mm.

After the routine CT scan, administer a contrast-enhanced scan using a dual-barrel high-pressure injector. Inject the iodinated contrast agent iohexol (350 mg/ml) at a dosage of 450 mg/kg of body weight via the patient’s right elbow vein, at a flow rate of 3-5 mL/s. Subsequently, inject 35mL of physiological saline. Employ automatic monitoring intelligent triggering technology. If the CT value of the patient’s abdominal aorta exceeds 120Hu, introduce a delay of 6 s and audibly prompt for inhalation and breath holding to perform arterial phase scanning. Then, delay for 60–70 s and 4-5 min to conduct venous phase and excretory phase scanning, respectively.

Following this, two radiologists should independently review the images. In case of any disagreements, a third physician should be consulted to reach a decision. Consensus agreement among the three radiologists forms the basis for all evaluations of imaging features. The radiologist measures the maximum diameter of retroperitoneal masses based on the largest axial dimension.

Texture analysis for radiomics features

The acquired CT images were firstly resampled with a voxel size of 1 × 1 × 1 mm. The voxel intensity values were discretized using a fixed bin width of 5 HU, and the images were further normalized to scale the signal intensity using fixed window level of 40 HU and window width of 500 HU to a range of 1–255 HU. The grayscale values of the images were standardized using Z score normalization. Finally, non-rigid registration of the CT images from different scanners was performed using SimpleITK to mitigate the effects of variations in scanning protocols or equipment on radiomic features. Three dimensional regions of interest (ROIs) delineation around each residual mass on all slices was performed using ITK-SNAP software (version 3.8.0) by a radiologist (H.L., 10 years of experience in radiology) after the residual nodal lesion was identified on the template RPLND pathology report and correlated with imaging. To ensure the reliability and reproducibility, each ROI was re-outlined by another radiologist (L.T., 15 years of experience in abdominal imaging diagnosis). To analyze the intraobserver agreement, the radiologist (H.L.) repeated the segmentation after the first delineation. To ensure greater precision in ROI drawing and given the challenges associated with automatic or semi-automatic segmentation of retroperitoneal soft tissues, we opted for a fully manual annotation approach. The manual delineation process involves selecting the most characteristic cross-section of the tumor and manually outlining it as closely as possible to the tumor’s boundary. After outlining this cross-section, we manually extend the delineation both upwards and downwards, layer by layer, until the tumor is no longer visible in the cross-section. The software then fills in the outlined regions to generate the three dimensional mask ROI for the target tumor.

Pyradiomics package (version 3.0.1) was utilized for radiomics features extracting, set to default to align with the Image Biomarker Standardization Initiative (IBSI) guidelines (https://pyradiomics.readthedocs.io/en/latest/faq.html).⁴² This process ensured a total of 1130 radiomics features were consistently extracted from the defined regions of interest (ROIs) on portal venous phase CT images. From the original images, 107 were extracted, including 32 first-order features (18 intensity statistical and 14 shape features). Among the 75 textural features, there were 24 Gy-level co-occurrence matrices (GLCM), 16 Gy-level run-length matrices (GLRLM), 16 Gy-level size zone matrices (GLSZM), 14 Gy-level dependence matrices (GLDM) and five neighboring gray-tone difference matrices (NGTDM) features. In addition, the same number of first-order grayscale statistical features and texture features were extracted based on different transformed images. A total of 744 features were extracted based on wavelet decomposition images with eight filter channels, and 279 features were extracted based on Laplacian of Gaussian (LoG) transform images (sigma parameters selected as 1.0 mm, 3.0 mm, and 5.0 mm).⁴² The features data extracted from one of the delineated ROIs and the intra-/interobserver correlation coefficients (ICCs) statistics are presented in the Table S10.

Feature selection and model establishment

For radiomics feature selection and model development, we used data from the training set. The process involved a four-step feature reduction. First, the intra- and interobserver ICCs were calculated and the features with intra- and interobserver ICCs values simultaneously greater than 0.75 were retained for further screening. Second, a univariate association test identified variables associated with histopathological outcomes, and only variables showing association (Student’s T-test, Welch’s T-test, and Mann-Whitney U-test for comparisons between two groups, ANOVA, Welch’s ANOVA, and Kruskal-Wallis H-test for comparisons between three groups, alpha = 0.05) were retained in the analysis. Third, after normalizing all numeric features using Z score standardization, pairwise correlation analysis eliminated highly correlated (Pearson correlation ≥0.8) redundant features. Finally, the least absolute shrinkage and selection operator (LASSO) regression with 10-fold cross-validation was performed. Features with non-zero coefficients in the LASSO analysis were used to build the radiomics signature and calculate the radiomics score (Radscore) for each patient (Radscore = intercept + feature 1 × coefficient 1 + feature 2 × coefficient 2 …). Additionally, to verify the discriminative power of radiomics, patients were divided into “benign” (necrosis/fibrosis) and “malignant” (teratoma or viable GCT). Similarly, the described procedure was used as screening features to distinguish necrosis/fibrosis from other histopathological components.

Supervised learning used six classifiers, such as regularized logistic regression (LR), support vector machine (SVM), random forest (RF), adaptive boosting (ADA), gradient boosting decision tree (GBDT), eXtreme Gradient Boosting (XGBoost), to identify a classifier that could correctly classify new objects.⁴⁵ The tuned model then was then applied to a randomly split validation set that had not been shown to the algorithm before. Bayesian optimization or grid search was used for parameter tuning to optimize the model’s performance, resulting in the optimal parameter combinations detailed in the Table S11. The optimal model was selected and adjusted based on its performance on the validation set, and finally used for internal prospective testing to evaluate its performance and generalizability. We also ensured baseline data consistency across randomly assigned cohorts and prospective test set to control for confounding factors.

RNA extraction and real-time PCR analysis

RNA was isolated from antecubital vein serum of NSGCT patients before pcRPLND and ten age-matched healthy male blood donors. In detail, during the week preceding the surgery, fasting venous blood (10 mL) was obtained from the antecubital vein of patients lying supine in the ward between 7 and 9 a.m. using standard venepuncture techniques. The whole blood was collected into in cell-free DNA BCT tubes (Streck, Inc.), stored at 4°C–8°C, and transported to clinical laboratories within 6h for processing. In order to obtain plasma, the blood samples were centrifuged at 3500 × g for 10 min at 4°C to harvest serum. The supernatant was then aliquoted and frozen at −80°C for storage. RNA was isolated from up to 200 μL of patient cubital vein serum by Trizol RNA extraction (Invitrogen) according to the manufacturer’s instructions. Final volume was brought up to 200 μL with sterile phosphate buffered saline for samples with lesser volumes prior to extraction. RNA was eluted into 100 μL of nuclease-free water and stored at −80°C until point of use.

The initial quality control (QC) assessment followed the protocol outlined by Murray et al.⁴⁶ cDNA synthesis and quantification were carried out using TaqMan MicroRNA assays in singleplex format (assay IDs listed below). QC procedures included evaluating hemolysis, extraction efficiency, and sample integrity. Any sample that failed the extraction efficiency or sample integrity tests was excluded, and another serum aliquot was processed for extraction. Nine samples did not pass the hemolysis test and were excluded. The remaining samples were included in the final analysis.

A multiplex RT and pre-amplification strategy was used, as described previously.⁴⁶ Four TaqMan primers were included in multiplex in this mixture: hsa-miR-371a-3p (assay ID: 002124), hsa-miR-375-5p (assay ID: 464184), cel-miR-39-3p (assay ID: 000200), and hsa-miR-30b-5p (assay ID: 000602) (Ruibiotech, China) (Table S12). The resulting cDNA was pre-amplified with the TaqMan Pre-Amp Master Mix, and diluted 5x prior to use. The pre-amplified cDNA was then analyzed by Mir-X miRNA First Strand Synthesis Kit (Takara, Japan) and Mir-X miRNA qRT-PCR SYBR Kit (Takara, Japan) was used according to the manufacture’s protocol. And qRT-PCR was carried out in 96-well plates using QuantstudioTM 7 Flex Real-Time PCR (Applied BiosystemsTM, Foster City, CA, USA), according to the recommended protocol, starting at 95°C for 20 s, followed by 40 cycles at 95°C for 15 s, 50°C for 30 s and 72°C for 30 s and terminated at 72°C for 15 s. Cycle threshold values were normalized to cel-miR-39-3p as an exogenous spike-in control and miR-30b-5p as an endogenous control. Target miRNA level was calculated relative to the mean expression of the miRNA in ten extracts from serum collected from healthy male donors. All experiments were done in triplicates. miRNA levels were determined according to the 2-ΔΔCt method. Thresholds were determined with Youden’s J statistic.

Quantification and statistical analysis

Statistical analysis

Statistical analyses were conducted using Python (version 3.7.4) and R software (version 3.6.1). Multiple imputation was used to fill in the missing values. Adaptive Synthetic Sampling Approach for Imbalanced Learning (ADASYN) was used exclusively for adaptive synthetic oversampling only in the training set to balance the different classes.⁴⁷ We employed independent t-tests, one-way ANOVA, the non-parametric Mann-Whitney U-test, and Kruskal-Wallis H-test to test group differences. The least significant difference and Dunnett’s T3 multiple comparison corrections were applied for post-hoc analyses. Categorical parameters were compared using the Pearson chi-squared test or Fisher’s exact test when appropriate. A two-tailed p value of <0.05 was considered to indicate statistical significance.

A late-fusion approach was used to establish the radiomics-miRNA integrated model. In detail, the prediction probability of radiomics model is combined with miRNA data, and LR was conducted on these newly formed predictive factors to establish the integrated model. To evaluate the classification performance of multi-class models, we used a macro-average approach. This involved calculating metrics for each binary classification scenario within the multi-class identification and averaging the results. The multiclass macro-average area under the receiver operating characteristics curve (AUC) and its confidence interval assessed the discriminative power and diagnostic accuracy of the models, with precision, specificity, sensitivity, and F1 score also calculated. The receiver operating characteristics (ROC) curves were compared using the Delong test, with Bonferroni correction for multiple comparisons. Calibration curves, with accompanying metrics like the Brier score and log loss, evaluated the calibration of the models, along with the Hosmer-Lemeshow test. AIC (Akaike Information Criterion) is a measure of model fit. We calculated the AIC value using the formula: $\text{AIC}=-2\ln \left(L\right)+2\mathrm{k}$, where k is the number of parameters in the fitted model, and L represents the log likelihood value, which indicates the fit between the model and the data. When the errors of the model follow an independent normal distribution, L is equivalent to the mean squared error. Decision curve analysis quantified the clinical utility of the models. A radiomics-miRNA nomogram, incorporating the radiomics scores (Radscores) and serum miRNA levels, was also developed. Finally, a web version of the dynamic nomogram APP was created using the “shinyPredict” packages.^34,35

The following Python (Version 3.7.4) packages were used: “os”, “pandas”, and “numpy” tidied the data or modified the format of the data; “scipy” was used for statistical analysis; “imbalanced-learn” implemented adaptive synthetic oversampling; “pyradiomics” extracted the radiomics features; “sklearn” implemented the dataset partitioning, LASSO regression, machine learning model construction, model evaluation, and calibration curve analysis; “bayes_opt” was used for Bayesian optimization; and “matplotlib.pyplot” and “seaborn” draws the graphs.

Simultaneously, the following R (version 4.1.0) packages were used: “mice” performed multiple imputation to fill in the missing values; “glmnet” calculated the radiomics score; “rmda” implemented the decision curve analysis; “regplot” constructed the nomogram; “shinyPredict” created the online prediction website; and “ggplot2” draws the graphs.

Additional resources

The website for the radiomics-miRNA integrated predictive system is as follows: https://dingrj.shinyapps.io/pcRPLND_prediction. The study has been registered on Chinese Clinical Trial Registry, the registration ID is ChiCTR2400081978. The access website is as follows: https://www.chictr.org.cn/bin/project/edit?pid=242861.

Published: December 6, 2024