Abstract
Purpose
This study aimed to evaluate CT- and MRI-based radiomic features of ovarian mature and immature teratomas (MTs and ITs) and to investigate the differences between low- and high-grade ITs.
Methods
A total of 41 patients aged ≤ 30 years with histopathologically confirmed ovarian MT and 23 patients with ovarian IT who underwent preoperative MRI were included. Among them, 51 (31 MTs and 20 ITs) underwent preoperative CT. We investigated the differences in visual findings by radiologists and radiomics analysis not only between MTs and ITs but also between low- (grade 1) and high-grade (grades 2–3) ITs.
Results
Multiple CT and MRI findings, along with radiomic features, showed significant differences between MTs and ITs (P < 0.05). Although the maximum diameter of nodular fat was greater in low-grade ITs than in high-grade ITs (P < 0.05), no other visual CT or MRI findings showed significant differences between low- and high-grade ITs. In contrast, several MRI-based radiomic features derived from fat-suppressed T1-weighted images, along with selected CT-based radiomic features, showed significant differences between low- and high-grade ITs (P < 0.05).
Conclusion
Visual assessments by radiologists and radiomics analysis effectively distinguished MTs from ITs. However, differentiating between low- and high-grade ITs using visual assessment alone was difficult. Radiomics analysis—particularly MRI-based features from fat-suppressed T1-weighted images—showed potential as a noninvasive approach for grading ITs, although further validation is required.
Keywords: ovary, teratoma, computed tomography, magnetic resonance imaging, radiomics
Introduction
Ovarian teratomas are the predominant histological subtype of ovarian germ cell tumors1 and are histologically categorized into mature and immature teratomas (MTs and ITs) based on the presence of immature neuroectodermal tissue within the tumors.2 Although fertility-sparing surgery is recommended for pediatric patients with both MT and IT,3,4 minimally invasive laparoscopic cystectomy is frequently considered the gold standard for MT, whereas unilateral salpingo-oophorectomy is the current standard for early-stage IT.5–7 According to the National Comprehensive Cancer Network guidelines, adjuvant chemotherapy is usually not required for stage I, grade 1 IT because of low recurrence risk after complete resection. However, stages I (grades 2 and 3) and II–IV ITs require postoperative chemotherapy with 3–4 cycles of bleomycin/etoposide/cisplatin.8 Moreover, recurrence after chemotherapy is more common in high-grade ITs than in low-grade ITs, and histological grade represents the most significant risk factor for recurrence.9 Therefore, accurate preoperative differentiation between low- and high-grade ITs on radiological imaging would be valuable for early treatment planning and risk stratification for recurrence and prognosis.
Imaging modalities such as ultrasound, CT, and MRI play an essential role in the preoperative evaluation of ovarian teratomas.10,11 CT effectively detects fat and calcification, whereas MRI offers excellent soft tissue contrast and can differentiate various tumor components. Previous studies have suggested that imaging features, including a larger tumor size, predominance of solid components, irregular margins, and the presence of scattered fatty or calcified areas, were associated with IT.10–15 For example, on CT, irregular margins (P = 0.005) and solid soft tissue lesions (P = 0.001) were observed more frequently in ITs, whereas mural nodules containing soft tissue, fat, or calcifications were characteristic findings of MTs (P = 0.044).11 On MRI, ITs appeared as fat-containing tumors with solid components composed of numerous cysts of varying sizes (100%), with signals predominantly similar to simple fluid in most cases (90%).12 However, to the best of our knowledge, no previous studies have comprehensively compared CT and MRI findings between low- and high-grade ITs.
Radiomics is an emerging technique that extracts high-dimensional, quantitative features from medical images, enabling comprehensive tumor characterization beyond human perception. Increasing evidence suggests that it may help differentiate between benign and malignant tumors, predict histological subtypes, and assess prognosis.16 Although CT-based texture analysis has been explored to differentiate MT from IT—for example, gray level co-occurrence matrix (GLCM) features show higher correlation, entropy, and dissimilarity in ITs, whereas MTs exhibit higher homogeneity and energy (P < 0.05)17—the utility of MRI-based radiomics analysis for distinguishing both ITs from MTs and high-grade from low-grade ITs has not yet been investigated.
Therefore, this study aimed (1) to assess the differences in CT and MRI findings through radiologist assessment and radiomic feature extraction between MTs and ITs, as well as between low- and high-grade ITs, and (2) to evaluate the diagnostic performance of radiomic features for differentiating these subtypes. This study aimed to offer insights into the potential role of radiomics in the noninvasive diagnostic evaluation of ovarian teratomas.
Materials and Methods
Patients
The present study was approved by the institutional review board of Gifu University and Central Japan International Medical Center and complied with the guidelines of the Health Insurance Portability and Accountability Act of 1996 and the Declaration of Helsinki. The requirement for informed consent was waived due to the retrospective nature of this study. Using the electronic medical record system of Gifu University and Central Japan International Medical Center, we identified patients with histopathologically confirmed ovarian MT and IT who underwent surgical resection between August 2006 and September 2024. The inclusion criteria were (a) preoperative MRI and (b) presence of a primary tumor, whereas the exclusion criteria were (a) patients with MT aged > 30 years, (b) inadequate image quality preventing radiological assessment, and (c) recurrent tumors. Due to exclusion criteria (a), patients with MT aged > 30 years were excluded to reduce age-related bias.
MRI protocols
All 64 patients underwent MRI examinations using 1.5T MRI scanners (Intera Achieva 1.5T Pulsar; Philips Healthcare, Best, the Netherlands, or Ingenia Prodiva 1.5T CS; Philips Healthcare) or a 3T MRI scanner (Intera Achieva 3.0T Quasar Dual; Philips Healthcare). Axial and sagittal non-fat-suppressed T2-weighted fast spin-echo (T2WI; TR/TE, 3000–7402/90–100 ms), axial non-fat-suppressed T1-weighted spin-echo (T1WI; TR/TE, 486–789/10–17 ms), and axial fat-suppressed T1-weighted spin-echo (FS-T1WI; TR/TE, 544–817/10–17 ms) images were obtained from all patients. MRI images were obtained at a section thickness of 4–8 mm with an intersection gap of 1–2 mm. Contrast-enhanced MRI images were not assessed in this study.
CT protocols
CT examinations were performed in 51 patients (31 MTs and 20 ITs) using a 16-slice CT scanner (LightSpeed 16; GE Healthcare, Milwaukee, WI, USA), or 64-slice CT scanners (Brilliance 64; Philips Healthcare, or Discovery CT750 HD; GE Healthcare). Axial unenhanced CT images were obtained at a section thickness of 5 mm without an intersection gap, and coronal reformatted images were reconstructed at a section thickness of 2.5 mm without an intersection gap. Contrast-enhanced CT images were not assessed in this study.
Visual assessment by radiologists
CT and MRI images were independently reviewed by 2 radiologists with 26 and 4 years of experience in gynecological imaging, respectively. Both were blinded to clinical and histopathological data. Disagreements were resolved through discussion by reaching a consensus. If the teratomas involved bilateral ovaries, the reviewers evaluated the imaging findings of the larger lesion.
Qualitative MRI assessment included the evaluation of the predominance of solid or cystic components, presence of solid and cystic components, loculation (uni- or multi-locular), tumor margin (well- or ill-defined), lobulated margin, septal thickening (> 2 mm), macroscopic fat, multiple fat foci, scattered fat, and characteristic MRI signs such as stained glass appearance, honeycomb sign, and fat-fluid level. Qualitative CT assessment included macroscopic fat (nodular, plate-like, multiple, and scattered) and calcification (nodular, plate-like, multiple, and scattered). Scattered fat and calcification were defined as elements dispersed throughout the tumor, varying in size, shape, and density. The honeycomb sign was defined as a densely aggregated cluster of numerous loculi measuring 5–10 mm. Nodular fat and calcification were defined as convex, protruding nodular structures. Plate-like fat and calcification were defined as thin, layered structures along the cyst wall or septum, rather than nodular.
Quantitative measurements included the maximum tumor diameter, the largest diameter of nodular fat on MRI, and the largest diameter of nodular calcification on CT.
Radiomics analysis
Radiomics analysis was performed for solid components on the slice depicting the largest solid component. The reviewer with 4 years of experience, blinded to the clinical and histopathological data, manually delineated the ROI to encompass the entire solid component according to the following unified criteria (Fig. 1). Accurately delineating the solid components using unenhanced CT and MRI alone was challenging due to the frequent presence of microcysts, particularly in ITs. Therefore, the solid components were defined as areas excluding regions ≥ 10 mm in size that exhibited uniformly low attenuation (0–20 Hounsfield units) on CT images or marked hyperintensity on T2WIs. If the teratomas involved bilateral ovaries, the reviewer placed the ROI on the larger lesion. For MRI, pixel intensities were normalized using the mean and standard deviation. No pixel resampling was applied. Feature extraction was conducted using PyRadiomics version 3.1.0.18 in Python 3.9.20 with a fixed gray-level discretization (binWidth = 25). Extracted features included (1) first-order statistics representing the distribution of voxel intensities, (2) 2D shape features, and (3) texture features such as entropy, autocorrelation, and cluster shade. Although various advanced texture analysis methods exist, we focused on the GLCM features because of their wide acceptance and representativeness in texture evaluation, aiming to gain an overview of potential differences in texture characteristics between histological subtypes. Radiomic features were extracted exclusively from the original images, without the use of any filtered or transformed images. Additional statistical analyses were conducted using the SciPy library (version 1.13.1), and data visualizations were performed using Matplotlib and Seaborn. The analysis was performed only in cases with available imaging data, and missing values were not imputed. Radiomic feature extraction and reporting were conducted with reference to the Imaging Biomarker Standardisation Initiative recommendations and the Checklist for the Evaluation of Radiomics research guideline.
Fig. 1.
Example of an image and mask used for radiomics analysis. The left panel shows a T2-weighted MRI image, and the right panel displays the corresponding manually created mask. Axial slices from MRI (T2-weighted, T1-weighted, and fat-suppressed T1-weighted images) and CT were reviewed to select the slice containing the largest extent of the tumor’s solid components. An ROI was manually delineated to fully encompass the solid components. Areas ≥ 10 mm in size that exhibited uniformly low attenuation (0–20 Hounsfield units) on CT images or marked hyperintensity on T2-weighted images were excluded. Using this mask, first-order statistics, shape, and texture features were extracted with the PyRadiomics library.
Feature-space analysis and robustness assessment
To visualize the multivariate structure of the radiomic feature space, correlation clustering and principal component analysis (PCA) were performed separately for T2WI, T1WI, FS-T1WI, and CT. Radiomic features were standardized across subjects using z-scores. Feature redundancy was assessed using pairwise correlation matrices, which were visualized as hierarchically clustered heatmaps (Supplementary Fig. 1). PCA was then applied to the same standardized feature matrix, and scatter plots of the first 2 principal components (PC1 and PC2) were generated to compare (i) MT vs. IT and (ii) low- vs. high-grade IT (Supplementary Fig. 2). Group-level distributional shift in the PCA space was quantified as the Euclidean distance between group centroids in the PC1–PC2 plane, and statistical significance was evaluated using a label-permutation test (Supplementary Table 1).
To assess feature robustness to segmentation variability, a perturbation-based reproducibility analysis was conducted. For each case, the original 2D ROI was morphologically eroded and dilated by ±1 pixel, and radiomic features were re-extracted. Reproducibility across the 3 ROI conditions (original, erosion, and dilation) was quantified using the intraclass correlation coefficient (ICC) based on a 2-way random-effects model with absolute agreement and single measurement.
Comparative analysis of low- and high-grade ITs
All features described above—including both visual assessments (qualitative and quantitative assessments) and radiomics analysis—were compared between low- and high-grade ITs, as well as between MTs and ITs.
We adopted a 2-tiered classification system to compare imaging findings between low-grade (grade 1) and high-grade (grades 2–3) ITs. This classification was chosen for several reasons: (1) treatment strategies differ for International Federation of Gynecology and Obstetrics stage I based on histological grade; (2) the small number of cases of grades 2 and 3 necessitated grouping for statistical robustness; (3) the 2-tiered system has better histopathological reproducibility19; and (4) it is recognized in international guidelines, including the WHO classification.20
Statistical analysis
Statistical analyses were performed using IBM SPSS Statistics version 24.0 (IBM, Armonk, NY, USA) or MedCalc Statistical Software (MedCalc for Windows, version 22.009; Mariakerke, Ostend, Belgium). Fisher’s exact test was performed to compare qualitative assessments between the 2 groups. Interobserver variability in qualitative assessments was evaluated using kappa (κ) statistics, where a κ value of ≤ 0.20 was interpreted as slight agreement, 0.21–0.40 as fair agreement, 0.41–0.60 as moderate agreement, 0.61–0.80 as substantial agreement, and ≥ 0.81 as almost perfect agreement. The Mann–Whitney U test was used to compare quantitative measurements between the 2 groups. Receiver operating characteristic (ROC) curve analysis was used to determine the diagnostic performance of selected radiomic features, and the area under the curve (AUC) was calculated. Interobserver variability in quantitative parameters and radiomic features was assessed using the ICC, where ICC of < 0.50 was interpreted as poor agreement, 0.50–0.75 as moderate agreement, 0.75–0.90 as good agreement, and > 0.90 as excellent agreement. P-values < 0.05 were considered statistically significant.
Model development and internal validation
The dataset was split into a training set and a hold-out validation set (7:3). Only features with ICC > 0.70 were included for analysis. Using the training set only, zero-variance features were removed and highly correlated features (Spearman’s |ρ| > 0.90) were reduced by retaining the feature with the higher univariate AUC. Prediction models were constructed using regularized logistic regression with least absolute shrinkage and selection operator (LASSO) (L1) and Elastic Net (L1/L2) penalties, and hyperparameters (regularization parameter λ and, for Elastic Net, the mixing parameter α) were optimized by internal cross-validation within the training set. Model performance was evaluated on the hold-out validation set using ROC analysis. For internal validation, 200-times stratified bootstrap resampling of the training set was performed. The mean AUC and the 95% confidence interval (CI) were obtained by refitting the model with re-optimized hyperparameters and calculating the AUC on the fixed validation set for each bootstrap iteration.
Results
Patient characteristics
Table 1 summarizes the patients’ ages and serum tumor marker levels.
Table 1.
Comparison of patients’ ages and tumor markers
| Parameter | MTs | ITs | P-value(AUC) | Low-grade ITs(grade 1) | High-grade ITs(grades 2–3) | P-value(AUC) |
|---|---|---|---|---|---|---|
| Age (years) | 24.1 ± 5.2 | 20.6 ± 7.0 | 0.029* | 20.9 ± 7.3 | 20.1 ± 6.8 | 0.950 |
| 26.0 (21.0–28.0) | 20.0 (17.0–24.0) | (0.665) | 21.0 (18.0–24.0) | 20.0 (16.0–24.0) | (0.508) | |
| n = 41 | n = 23 | n = 14 | n = 9 | |||
| AFP (ng/mL) | 2.9 ± 1.2 | 148.0 ± 312.6 | 0.001* | 180.1 ± 412.3 | 105.2 ± 110.7 | 0.651 |
| 2.8 (2.3–3.4) | 44.8 (7.6–116.3) | (0.883*) | 35.3 (14.4–65.7) | 89.4 (16.2–147.1) | (0.573) | |
| n = 14 | n = 14 | n = 8 | n = 6 | |||
| CA125 (U/mL) | 29.3 ± 21.9 | 74.6 ± 65.9 | 0.005* | 78.7 ± 78.5 | 70.5 ± 54.9 | 0.895 |
| 21.1 (14.1–35.3) | 60.0 (27.9–85.1) | (0.743) | 64.0 (17.5–71.5) | 44.3 (29.1–88.2) | (0.519) | |
| n = 32 | n = 18 | n = 9 | n = 9 | |||
| CA19-9 (U/mL) | 594.0 ± 1388.7 | 221.1 ± 376.2 | 0.077 | 151.3 ± 272.9 | 283.3 ± 456.7 | 0.124 |
| 49.3 (10.7–237.8) | 50.8 (25.2–122.2) | (0.526) | 26.8 (20.2–102.2) | 86.3 (50.8–122.2) | (0.722) | |
| n = 33 | n = 17 | n = 8 | n = 9 | |||
| SCC (ng/mL) | 1.6 ± 0.8 | 2.1 ± 1.0 | 0.113 | 2.4 ± 0.9 | 2.0 ± 1.2 | 0.327 |
| 1.4 (1.0–2.0) | 1.9 (1.4–2.6) | (0.680) | 2.2 (1.8–2.8) | 1.4 (1.3–2.3) | (0.700) | |
| n = 25 | n = 9 | n = 4 | n = 5 | |||
| CEA (ng/mL) | 2.0 ± 1.4 | 1.2 ± 0.5 | 0.223 | 1.2 ± 0.4 | 1.2 ± 0.6 | 0.639 |
| 1.4 (1.2–2.1) | 1.2 (0.9–1.6) | (0.702) | 1.2 (1.1–1.4) | 1.4 (0.9–1.7) | (0.625) | |
| n = 7 | n = 6 | n = 2 | n = 4 |
Data are presented as mean ± standard deviation, median (interquartile range), followed by the sample size (n). P-values with asterisks (*) indicate statistical significance (P < 0.05) or high diagnostic performance (AUC > 0.80).
AFP, alpha-fetoprotein; AUC, area under the curve; CA, cancer antigen; CA19-9, carbohydrate antigen 19-9; CEA, carcinoembryonic antigen; IT, immature teratoma; MT, mature teratoma; SCC, squamous cell carcinoma antigen.
A total of 64 patients were included in this analysis, comprising 41 patients aged ≤ 30 years with MT and 23 patients with IT who had preoperative MRI. The number of MT cases was relatively low because the study included only patients ≤ 30 years of age to minimize age-related bias and because the analysis was restricted to cases in which tumor resection had been performed. MTs were bilateral in 7 patients and unilateral in the remaining 34, whereas ITs were unilateral in all 23 patients. The mean age was lower in the IT group (20.6 years) than in the MT group (24.1 years) (P = 0.029). However, no significant age difference was observed between patients with low- and high-grade ITs (P = 0.950).
Serum alpha-fetoprotein (AFP) (P = 0.001) and cancer antigen 125 (CA125) (P = 0.005) levels were higher in the IT group than in the MT group. In contrast, no significant differences were observed in serum carbohydrate antigen 19-9 (CA19-9) (P = 0.077), squamous cell carcinoma antigen (SCC) (P = 0.113), and carcinoembryonic antigen (CEA) (P = 0.223). Meanwhile, no significant differences in tumor markers were found between low- and high-grade ITs: AFP (P = 0.651), CA125 (P = 0.895), CA19-9 (P = 0.124), SCC (P = 0.327), and CEA (P = 0.639).
CT and MRI findings
MT vs. IT
Table 2 summarizes the qualitative findings of CT and MRI. On MRI, the following features were more frequent in ITs than in MTs: solid components (96% vs. 41%, P < 0.001), lobulated morphology (61% vs. 32%, P = 0.035), multiple fat foci (95% vs. 70%, P = 0.023), scattered fat (95% vs. 8%, P < 0.001), stained glass appearance (96% vs. 56%, P = 0.001), and honeycomb sign (100% vs. 10%, P < 0.001). Conversely, no significant differences were observed in multilocularity (P = 0.150), irregular surface (P = 0.359), macroscopic fat (P = 1.000), and fat-fluid level (P = 0.241). On CT, ITs more frequently showed plate-like fat (90% vs. 61%, P = 0.029), scattered fat (90% vs. 7%, P < 0.001), nodular calcification (100% vs. 77%, P = 0.034), plate-like calcification (90% vs. 23%, P < 0.001), multiple foci of calcification (100% vs. 52%, P < 0.001), and scattered calcification (100% vs. 8%, P < 0.001). No significant differences were observed in the presence of any fat (P = 1.000), nodular fat (P = 0.636), multiple fat foci (P = 0.163), and any calcification (P = 0.070).
Table 2.
Comparison of qualitative imaging features based on visual assessments by radiologists
| Imaging features | MTs | ITs | P-value(AUC) | Low-grade ITs(grade 1) | High-grade ITs(grades 2–3) | P-value(AUC) | κ-value |
|---|---|---|---|---|---|---|---|
| MRI features | n = 41 | n = 23 | n = 14 | n = 9 | |||
| Predominant component – cystic | 41/41 (100%) | 21/23 (91%) | 0.125 | 14/14 (100%) | 7/9 (78%) | 0.142 | 0.760 |
| (0.543) | (0.611) | ||||||
| Solid component | 17/41 (41%) | 22/23 (96%) | < 0.001* | 14/14 (100%) | 8/9 (89%) | 0.391 | 0.522 |
| (0.771) | (0.556) | ||||||
| Cystic component | 41/41 (100%) | 23/23 (100%) | 1.000 | 14/14 (100%) | 9/9 (100%) | 1.000 | NA |
| (0.500) | (0.500) | ||||||
| Multilocularity | 36/41 (88%) | 23/23 (100%) | 0.150 | 14/14 (100%) | 9/9 (100%) | 1.000 | 0.524 |
| (0.561) | (0.500) | ||||||
| Irregular surface | 0/41 (0%) | 1/23 (4%) | 0.359 | 0/14 (0%) | 1/9 (11%) | 0.391 | NA |
| (0.522) | (0.556) | ||||||
| Lobulated morphology | 13/41 (32%) | 14/23 (61%) | 0.035* | 8/14 (57%) | 6/9 (67%) | 1.000 | 0.450 |
| (0.646) | (0.548) | ||||||
| Septal thickening | 35/41 (85%) | 23/23 (100%) | 0.080 | 14/14 (100%) | 9/9 (100%) | 1.000 | 0.436 |
| (0.573) | (0.500) | ||||||
| Macroscopic fat | 40/41 (98%) | 22/23 (96%) | 1.000 | 14/14 (100%) | 8/9 (89%) | 0.391 | 0.760 |
| (0.510) | (0.556) | ||||||
| Multiple fat foci | 28/40 (70%) | 21/22 (95%) | 0.023* | 14/14 (100%) | 7/8 (88%) | 0.364 | 0.481 |
| (MTs: n = 40, ITs: n = 22) | |||||||
| (0.627) | (0.562) | ||||||
| Scattered fat | 3/40 (8%) | 21/22 (95%) | < 0.001* | 14/14 (100%) | 7/8 (88%) | 0.364 | 0.562 |
| (MTs: n = 40, ITs: n = 22) | (0.940*) | ||||||
| (0.562) | |||||||
| Stained glass appearance | 23/41 (56%) | 22/23 (96%) | 0.001* | 14/14 (100%) | 8/9 (89%) | 0.391 | 0.635 |
| (0.698) | (0.556) | ||||||
| Honeycomb sign | 4/41 (10%) | 23/23 (100%) | < 0.001* | 14/14 (100%) | 9/9 (100%) | 1.000 | 0.902 |
| (0.951*) | (0.500) | ||||||
| Fat-fluid level | 7/41 (17%) | 1/23 (4%) | 0.241 | 0/14 (0%) | 1/9 (11%) | 0.391 | 0.572 |
| (0.564) | |||||||
| (0.556) | |||||||
| CT features | n = 31 | n = 20 | n = 12 | n = 8 | |||
| Fat presence | 30/31 (97%) | 20/20 (100%) | 1.000 | 12/12 (100%) | 8/8 (100%) | 1.000 | 0.758 |
| (0.516) | (0.500) | ||||||
| Nodular fat | 27/31 (87%) | 19/20 (95%) | 0.636 | 12/12 (100%) | 7/8 (88%) | 0.400 | 0.657 |
| (0.540) | (0.562) | ||||||
| Plate-like fat | 19/31 (61%) | 18/20 (90%) | 0.029* | 10/12 (83%) | 8/8 (100%) | 0.495 | 0.435 |
| (0.644) | (0.583) | ||||||
| Multiple fat foci | 21/30 (70%) | 18/20 (90%) | 0.163 | 11/12 (92%) | 7/8 (88%) | 1.000 | 0.750 |
| (0.600) | (0.521) | ||||||
| Scattered fat | 2/30 (7%) | 18/20 (90%) | < 0.001* | 11/12 (92%) | 7/8 (88%) | 1.000 | 0.480 |
| (0.917*) | (0.521) | ||||||
| Calcification presence | 25/31 (81%) | 20/20 (100%) | 0.070 | 12/12 (100%) | 8/8 (100%) | 1.000 | 0.867 |
| (0.597) | (0.500) | ||||||
| Nodular calcification | 24/31 (77%) | 20/20 (100%) | 0.034* | 12/12 (100%) | 8/8 (100%) | 1.000 | 0.833 |
| (0.613) | (0.500) | ||||||
| Plate-like calcification | 7/31 (23%) | 18/20 (90%) | < 0.001* | 11/12 (92%) | 7/8 (88%) | 1.000 | 0.826 |
| (0.837*) | (0.521) | ||||||
| Multiple foci of calcification | 13/25 (52%) | 20/20 (100%) | < 0.001* | 12/12 (100%) | 8/8 (100%) | 1.000 | 0.622 |
| (0.740) | (0.500) | ||||||
| Scattered calcification | 2/25 (8%) | 20/20 (100%) | < 0.001* | 12/12 (100%) | 8/8 (100%) | 1.000 | 0.461 |
| (0.960*) | (0.500) |
Asterisks (*) indicate statistical significance (P < 0.05) or high diagnostic performance (AUC > 0.80).
AUC, area under the curve; IT, immature teratoma; κ, kappa; MT, mature teratoma; NA, not applicable.
Table 3 presents the quantitative measurements. The maximum tumor diameter was greater in ITs than in MTs (168.4 ± 40.0 mm vs. 95.2 ± 47.1 mm, P < 0.001), whereas the maximum diameter of nodular fat was greater in MTs than in ITs (50.8 ± 24.9 mm vs. 16.1 ± 8.7 mm, P < 0.001). No significant difference was observed in the maximum diameter of nodular calcifications (P = 0.078).
Table 3.
Comparison of quantitative measurements by radiologist
| Parameter | MTs | ITs | P-value(AUC) | Low-grade ITs(grade 1) | High-grade ITs(grades 2–3) | P-value(AUC) | ICC |
|---|---|---|---|---|---|---|---|
| Max tumor diameter (mm) | 95.2 ± 47.1 | 168.4 ± 40.0 | < 0.001* | 175.5 ± 47.0 | 157.2 ± 24.4 | 0.395 | 0.864 |
| (33–215) | (93–270) | (0.881*) | (93–270) | (115–177) | (0.607) | ||
| n = 41 | n = 23 | n = 14 | n = 9 | ||||
| Max nodular fat diameter (mm) | 50.8 ± 24.9 | 16.1 ± 8.7 | < 0.001* | 19.3 ± 9.1 | 10.6 ± 4.2 | 0.022* | 0.832 |
| (8–100) | (3–36) | (0.903*) | (3–36) | (5–15) | (0.821*) | ||
| n = 27 | n = 19 | n = 12 | n = 7 | ||||
| Max nodular calcification diameter (mm) | 10.0 ± 5.6 | 13.1 ± 6.8 | 0.078 | 15.3 ± 7.8 | 10.0 ± 3.1 | 0.131 | 0.509 |
| (3–21) | (4–28) | (0.640) | (4–28) | (5–15) | (0.703) | ||
| n = 24 | n = 20 | n = 12 | n = 8 |
Data are presented as mean ± standard deviation (range), followed by the sample size (n). P-values with asterisks (*) indicate statistical significance (P < 0.05) or high diagnostic performance (AUC > 0.80).
AUC, area under the curve; ICC, intraclass correlation coefficient; IT, immature teratoma; MT, mature teratoma.
Low-grade vs. high-grade IT
The maximum diameter of nodular fat was larger in low-grade ITs than in high-grade ITs (19.3 ± 9.1 mm vs. 10.6 ± 4.2 mm, P = 0.022). However, no significant differences were observed in any qualitative imaging findings and most of the quantitative measurements between low- and high-grade ITs.
Interobserver agreement
The κ-values ranged from 0.435 to 0.902, indicating moderate to almost perfect agreement between the 2 radiologists. Interobserver agreement was good to excellent for features useful in differentiating IT from MT, including the honeycomb sign on MRI (κ = 0.902) and calcification-related findings on CT (κ = 0.826–0.867). In contrast, several morphological and fat-related features showed only moderate agreement (κ = 0.435–0.562). The ICCs ranged from 0.509 to 0.864, indicating moderate to good agreement between the 2 reviewers.
CT- and MRI-based radiomic features
MT vs. IT
Tables 4 and 5 summarize the radiomic features obtained from CT and MRI. Among first-order histogram features, CT-based radiomic features demonstrated the highest discriminative performance, with 15 of 18 features reaching statistical significance (P < 0.05). Notably, several CT-based radiomic features achieved AUC > 0.80, including: 10th percentile (AUC = 0.920, P < 0.001), median (AUC = 0.896, P < 0.001), robust mean absolute deviation (AUC = 0.885, P < 0.001), interquartile range (AUC = 0.879, P < 0.001), uniformity (AUC = 0.852, P < 0.001), mean (AUC = 0.847, P < 0.001), mean absolute deviation (AUC = 0.830, P < 0.001), and entropy (AUC = 0.812, P < 0.001).
Table 4.
Summary of radiomic features showing significant differences (P-value < 0.05) and high AUC (AUC > 0.80) for differentiating both IT from MT and high-grade from low-grade ITs
| MTs vs. ITs | Low-grade vs. high-grade ITs | |||
|---|---|---|---|---|
| P-value < 0.05 | AUC > 0.80 | P-value < 0.05 | AUC > 0.80 | |
| First-order features | ||||
| T2WIa | 9/16 | 3/16 | 0/16 | 0/16 |
| T1WIa | 9/16 | 4/16 | 1/16 | 0/16 |
| FS-T1WIa | 12/16 | 2/16 | 6/16 | 4/16 |
| CT | 15/18 | 8/18 | 2/18 | 0/18 |
| Shape features (2D) | ||||
| T2WI | 6/9 | 0/9 | 0/9 | 0/9 |
| T1WI | 5/9 | 0/9 | 0/9 | 0/9 |
| FS-T1WI | 6/9 | 0/9 | 0/9 | 0/9 |
| CT | 7/9 | 0/9 | 0/9 | 0/9 |
| Texture features (GLCM) | ||||
| T2WI | 20/24 | 1/24 | 0/24 | 0/24 |
| T1WI | 22/24 | 0/24 | 1/24 | 0/24 |
| FS-T1WI | 21/24 | 0/24 | 4/24 | 1/24 |
| CT | 10/24 | 1/24 | 1/24 | 0/24 |
The numbers are shown as n/N, where N is the total number of radiomic features extracted for each feature class and imaging modality (T2WI, T1WI, FS-T1WI, and CT), and n is the number of features meeting the criterion in each column (P-value < 0.05 or AUC > 0.80).
Due to z-score normalization, the mean and the variance (variance = standard deviation2) were excluded.
AUC, area under the curve; FS-T1WI, fat-suppressed T1-weighted image; GLCM, gray level co-occurrence matrix; IT, immature teratoma; MT, mature teratoma; T1WI, T1-weighted image; T2WI, T2-weighted image.
Table 5.
Statistical analysis results of first-order features for differentiating both IT from MT and high-grade from low-grade ITs
| MTs vs ITs | Low-grade vs high-grade ITs | |||||||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| T2WI | T1WI | FS-T1WI | CT | T2WI | T1WI | FS-T1WI | CT | |||||||||
| P-value | AUC | P-value | AUC | P-value | AUC | P-value | AUC | P-value | AUC | P-value | AUC | P-value | AUC | P-value | AUC | |
| 10th percentile | < 0.001* | 0.797 | 0.002* | 0.739 | 0.079 | 0.633 | < 0.001* | 0.920* | 0.900 | 0.516 | 0.044* | 0.754 | 0.038* | 0.762 | 0.342 | 0.631 |
| 90th percentile | < 0.001* | 0.846* | 0.861 | 0.513 | 0.005* | 0.712 | < 0.001* | 0.789 | 0.571 | 0.571 | 0.378 | 0.611 | 0.659 | 0.556 | 0.062 | 0.756 |
| Energy | 0.025* | 0.670 | 0.070 | 0.637 | 0.045* | 0.652 | 0.418 | 0.569 | 0.257 | 0.643 | 0.115 | 0.698 | 0.115 | 0.698 | 0.364 | 0.625 |
| Entropy | 0.005* | 0.714 | 0.861 | 0.513 | 0.232 | 0.591 | < 0.001* | 0.812* | 0.850 | 0.524 | 0.115 | 0.698 | 0.801 | 0.532 | 0.048* | 0.773 |
| Interquartile range | 0.534 | 0.547 | 0.002* | 0.736 | 0.049* | 0.649 | < 0.001* | 0.879* | 0.950 | 0.508 | 0.208 | 0.659 | 0.101 | 0.706 | 0.068 | 0.750 |
| Kurtosis | 0.872 | 0.512 | < 0.001* | 0.818* | 0.021* | 0.676 | 0.001* | 0.779 | 0.900 | 0.516 | 0.900 | 0.516 | 0.023* | 0.786 | 0.364 | 0.625 |
| Maximum | < 0.001* | 0.773 | < 0.001* | 0.925* | 0.016* | 0.682 | 0.027* | 0.688 | 0.801 | 0.532 | 0.186 | 0.667 | 0.005* | 0.857* | 0.248 | 0.659 |
| Mean absolute deviation | 0.152 | 0.609 | < 0.001* | 0.779 | 0.021* | 0.674 | < 0.001* | 0.830* | 0.900 | 0.516 | 0.571 | 0.571 | 0.078 | 0.722 | 0.099 | 0.727 |
| Meana | — | — | — | — | — | — | < 0.001* | 0.847* | — | — | — | — | — | — | 0.083 | 0.739 |
| Median | < 0.001* | 0.844* | 0.004* | 0.719 | 0.013* | 0.688 | < 0.001* | 0.896* | 0.413 | 0.603 | 0.089 | 0.714 | 0.008* | 0.833* | 0.214 | 0.670 |
| Minimum | 0.095 | 0.627 | 0.928 | 0.507 | < 0.001* | 0.830* | < 0.001* | 0.784 | 0.413 | 0.603 | 0.208 | 0.659 | 0.284 | 0.635 | 0.083 | 0.739 |
| Range | 0.186 | 0.600 | < 0.001* | 0.879* | < 0.001* | 0.835* | 0.396 | 0.572 | 1.000 | 0.500 | 0.186 | 0.667 | 0.008* | 0.833* | 0.215 | 0.670 |
| Robust mean absolute deviation | 0.345 | 0.572 | < 0.001* | 0.760 | 0.024* | 0.671 | < 0.001* | 0.885* | 0.950 | 0.508 | 0.231 | 0.651 | 0.166 | 0.675 | 0.058 | 0.761 |
| Root mean squared | 0.084 | 0.601 | 1.000 | 0.500 | 0.189 | 0.580 | 0.004* | 0.744 | 0.538 | 0.567 | 0.072 | 0.675 | 0.382 | 0.595 | 0.048* | 0.773 |
| Skewness | < 0.001* | 0.876* | < 0.001* | 0.835* | 0.016* | 0.683 | 0.008* | 0.725 | 0.614 | 0.563 | 0.488 | 0.587 | 0.004* | 0.865* | 0.364 | 0.625 |
| Total Energy | 0.025* | 0.670 | 0.070 | 0.637 | 0.045* | 0.652 | 0.418 | 0.569 | 0.257 | 0.643 | 0.115 | 0.698 | 0.115 | 0.698 | 0.364 | 0.625 |
| Uniformity | 0.005* | 0.714 | 0.861 | 0.513 | 0.232 | 0.591 | < 0.001* | 0.852* | 0.850 | 0.524 | 0.115 | 0.698 | 0.801 | 0.532 | 0.069 | 0.750 |
| Variancea | — | — | — | — | — | — | 0.002* | 0.767 | — | — | — | — | — | — | 0.186 | 0.682 |
The P-values and AUCs of radiomic features extracted from each imaging modality (CT, T2WI, T1WI, and FS-T1WI) are provided to evaluate their discriminative performance.
Due to z-score normalization, the mean and the variance (variance = standard deviation2) were excluded. Asterisks (*) indicate statistically significant P-values (P < 0.05) or high diagnostic performance (AUC > 0.80).
AUC, area under the curve; FS-T1WI, fat-suppressed T1-weighted image; IT, immature teratoma; MT, mature teratoma; T1WI, T1-weighted image; T2WI, T2-weighted image.
Radiomic features from FS-T1WI also showed significant differences, with 12 of 16 first-order features reaching statistical significance (P < 0.05). Regarding 2D shape-based features, a moderate number of significant differences were observed across modalities and sequences; however, none achieved an AUC of > 0.80, indicating limited discriminative power based on morphology alone. Furthermore, texture analysis revealed strong discriminatory ability, with 22, 21, and 20 of 24 GLCM features showing statistical significance in T1WI, FS-T1WI, and T2WI, respectively (P < 0.05).
Low-grade vs. high-grade IT
Among first-order features, FS-T1WI-based radiomic features showed the best discriminative performance, with 6 of 16 features reaching statistical significance. Notably, the following features had AUC > 0.80: skewness (AUC = 0.865, P = 0.004), maximum (AUC = 0.857, P = 0.005), median (AUC = 0.833, P = 0.008), and range (AUC = 0.833, P = 0.008) (Figs. 2 and 3). In contrast, CT-based first-order features showed limited discriminative ability, with only 2 of 18 being statistically significant (P < 0.05) and none exceeding an AUC of 0.80.
Fig. 2.
Comparison of representative first-order radiomic features extracted from FS-T1WI between low- and high-grade ITs. The left panels display violin plots with overlaid box plots and individual data points for skewness, maximum, median, and range, showing different distributions between low- and high-grade ITs (P < 0.01 for all). The right panels show the corresponding ROC curves, with AUC values ranging from 0.833 to 0.865, indicating high diagnostic performance in distinguishing between low- and high-grade ITs.AUC, area under the curve; FS-T1WI, fat-suppressed T1-weighted image; ITs, immature teratomas; ROC, receiver operating characteristic.
Fig. 3.
KDE plots of pixel intensity distributions from MRI (T2WI, T1WI, and FS-T1WI) and CT, comparing low- and high-grade ITs. For MRI, the distributions are displayed after z-score normalization. In all panels, the shaded areas indicate ±1 standard deviation (σ) from the mean. On FS-T1WIs, the peak of the high-grade IT is shifted more negatively compared to low-grade IT.FS-T1WI, fat-suppressed T1-weighted image; IT, immature teratoma; KDE, kernel density estimation; T1WI, T1-weighted image; T2WI, T2-weighted image.
No significant shape-based features were identified between low- and high-grade ITs (all AUCs < 0.80). Texture analysis revealed some statistically significant GLCM features, particularly in FS-T1WI (4 of 24), with only one feature exceeding an AUC of 0.80. Texture features derived from T1WI, T2WI, and CT provided limited differentiation between low- and high-grade ITs.
Feature-space and robustness analyses
Hierarchically clustered correlation heatmaps demonstrated clear patterns of feature redundancy, with multiple groups of highly correlated features forming consistent clusters (Supplementary Fig. 1). In PCA space, MT and IT exhibited a statistically significant centroid shift in the PC1–PC2 plane (Supplementary Fig. 2a and Supplementary Table 1). For IT grading, permutation testing in PCA space yielded the lowest P-value for FS-T1WI; however, this did not reach statistical significance (Supplementary Fig. 2b and Supplementary Table 1).
In the robustness analysis, key FS-T1WI first-order features used for grading in the primary analysis showed good-to-excellent reproducibility (skewness ICC = 0.960; maximum ICC = 0.876; median ICC = 0.866; range ICC = 0.887). Most features showed moderate to excellent reproducibility (ICC ≥ 0.50), while only a few features had ICC < 0.50 (e.g., T2WI 10th percentile ICC = 0.426; CT GLCM autocorrelation ICC = 0.416).
Internal validation of radiomics models
In MT vs. IT classification, the LASSO model achieved an AUC of 0.923 in the hold-out validation set, and the mean AUC in the bootstrap analysis was 0.845 (95% CI: 0.702–0.943). The Elastic Net model achieved an AUC of 0.942 in the hold-out validation set, and the mean AUC in the bootstrap analysis was 0.856 (95% CI: 0.721–0.952). In low-grade vs. high-grade IT classification, the LASSO model yielded an AUC of 0.562 in the hold-out validation set, and the mean AUC in the bootstrap analysis was 0.702 (95% CI: 0.434–1.000). The Elastic Net model yielded an AUC of 0.625 in the hold-out validation set, and the mean AUC in the bootstrap analysis was 0.717 (95% CI: 0.375–1.000) (Supplementary Fig. 3). Internal validation supported high discrimination performance between MTs and ITs. In contrast, the model performance for discriminating high-grade IT from low-grade IT was less stable, likely due to the limited sample size. The radiomic features retained in the final LASSO and Elastic Net models for each classification task are summarized in Supplementary Table 2.
Discussion
This study demonstrated the utility of visual assessment by radiologists in differentiating MTs from ITs. MRI revealed significant differences between MTs and ITs in features such as solid components, lobulated morphology, multiple fat foci, scattered fat, stained glass appearance, honeycomb sign, maximum tumor diameter, and maximum nodular fat diameter (Figs. 4, 5, 6). On CT, distinguishing features included plate-like fat, scattered fat, nodular calcification, plate-like calcification, multiple foci of calcification, and scattered calcification (Figs. 4, 5, 6). Therefore, radiologists can visually and reliably differentiate between MTs and ITs based on characteristic imaging features. These imaging differences are consistent with the histopathological characteristics of the tumors. MTs comprise well-differentiated tissues, including sebaceous glands and epithelial components that produce mucin. Because of the predominance of ectodermal components, such as skin, hair, and enamel, tooth and bone structures frequently form within the Rokitansky nodule. Consequently, abundant fat content and dense calcification are hallmark imaging features of MTs.10,14 In contrast, ITs typically contain solid components comprising immature neuroectodermal tissue, interspersed with cartilage, immature adipose tissue, and ossified foci. Although some mature components such as glands or fat may be present, their volume is usually limited. Therefore, imaging features such as solid components, scattered fat, and scattered calcification are more characteristic of ITs.10,12 In addition, ITs often exhibit microcystic structures lined with respiratory epithelium and filled with serous fluid, which form the “honeycomb sign” on imaging when clustered.21 The histological grading of ITs is determined by the amount of immature neuroepithelial tissue present.19 Among all imaging findings assessed by radiologists, only the maximum diameter of nodular fat differed between low- and high-grade ITs. Although all ITs showed scattered tiny hyperintense foci on FS-T1WI, their distribution and extent were variable and not clearly related to grade, making visual grading difficult. This indicates that visual assessment alone is insufficient for grading ITs, underscoring the need for more objective and quantitative evaluation methods.
Fig. 4.
A 23-year-old woman with bilateral ovarian mature teratomas. (a) T1-weighted, (b) T2-weighted, (c) fat-suppressed T1-weighted images, and (d) CT image. Cystic lesions containing abundant fat components are observed in both ovaries.
Fig. 5.
A 22-year-old woman with a right ovarian immature teratoma (low-grade, grade 1). (a) T1-weighted, (b) T2-weighted, (c) fat-suppressed T1-weighted images, and (d) CT image. Scattered fat and calcification, stained glass appearance, and honeycomb sign are observed.
Fig. 6.
A 23-year-old woman with a right ovarian immature teratoma (high-grade, grade 3). (a) T1-weighted, (b) T2-weighted, (c) fat-suppressed T1-weighted images, and (d) CT image. Predominantly solid components, scattered fat and calcification, stained glass appearance, and honeycomb sign are observed.
Serum tumor markers offer supplementary diagnostic value. Herein, AFP and CA125 levels were higher in ITs, whereas CA19-9 levels tended to be higher in MTs. These findings are consistent with previous research, showing elevated levels of CA125, CA153, and AFP in ITs compared with MTs.22 However, no tumor markers could reliably distinguish between low- and high-grade ITs, highlighting their limited role in grading.
Radiomics analysis provided further insights into differentiating between MTs and ITs and classifying ITs by histological grade. While comparing MTs and ITs, numerous radiomic features differed significantly, reflecting fundamental differences in their imaging characteristics. FS-T1WI-based first-order features demonstrated the strongest discriminative ability for distinguishing between low- and high-grade ITs (Fig. 2). Features such as skewness, maximum, median, and range were different, with high-grade ITs exhibiting greater skewness and a longer right-tailed distribution (Fig. 3). This pattern potentially results from scattered high-signal areas (e.g., hemorrhage, mucin, protein-rich components) within the tumor, which are accentuated by FS-T1WI. Furthermore, texture analysis using the GLCM revealed significant differences in several features between low- and high-grade ITs, including cluster shade, joint average, sum average, and autocorrelation on FS-T1WI; cluster tendency on T1WI; and informational measure of correlation 2 on unenhanced CT. These features quantify the distribution of pixel intensities and their spatial relationships and are sensitive to subtle variations in tumor signal distribution and heterogeneity. FS-T1WI potentially enhances the visibility of nonfat high-signal components, such as hemorrhage, mucin, and proteinaceous material, contributing to the differentiation of tumor grades.
Hierarchically clustered correlation analysis showed redundancy among features, suggesting that multiple significant features may reflect similar image characteristics. PCA revealed a centroid shift between MT and IT, indicating an overall difference in the radiomic feature space. For IT grading, FS-T1WI yielded the lowest permutation P-value, suggesting a trend toward separation, although this did not reach statistical significance. In addition, the ±1-pixel ROI perturbation analysis supported the robustness of the main findings to small segmentation changes.
The internal validation results suggested that radiomics-based models can reliably distinguish IT from MT. In both the LASSO and Elastic Net models, the bootstrap mean AUC was high, indicating that model performance was not strongly dependent on a specific training subset. In contrast, the models for discriminating high-grade from low-grade IT showed acceptable performance, but the results were less stable. This instability is likely explained by the limited sample size and the small validation set. Although these results support the feasibility of noninvasive grading, the current predictive models remain exploratory, and external validation in larger, independent multicenter cohorts is essential to confirm the clinical utility of CT- and MRI-based radiomics for IT grading.
In the final models, MT and IT could be differentiated using a relatively small number of first-order features, while differentiation between low- and high-grade ITs required a combination of multiple radiomic features. This suggests that the difference between MT and IT may be explained mainly by differences in the histogram-based signal-intensity distribution of the solid components of the tumor. For differentiating low- from high-grade ITs, first-order features derived from FS-T1WI and T1WI, such as skewness, maximum, and kurtosis, were retained in both models, suggesting that tumor grade may be associated with differences in the overall signal-intensity distribution as well as the presence of focal high-signal areas within the solid components. In addition, cluster tendency derived from T2WI was selected in both models, suggesting that the spatial pattern of signal heterogeneity may also contribute to grading. However, some of the selected features may represent correlated feature groups rather than completely independent indicators.
This study has several limitations that should be acknowledged. (1) This was a retrospective, two-center study with a relatively small sample size. (2) Images were obtained from various CT and MRI scanners, potentially introducing inter-scanner variability. (3) Contrast-enhanced CT and diffusion-weighted imaging were unavailable for some patients and thus were not assessed in this study. (4) Because detailed histopathologic correlation was not available for all cases, direct validation of the biologic basis of our findings was limited; future pathology studies incorporating digital pathology-based quantification would be valuable. (5) Manual tumor segmentation was performed, which may have introduced operator bias. In addition, radiomics analysis was performed using 2D segmentation, which may limit full tumor characterization, rather than volumetric 3D segmentation. (6) Multiplicity corrections (e.g., Bonferroni) were not applied because of the small sample size and the exploratory nature of the study. (7) The small sample size in this study limits the robustness and generalizability of predictive modeling, especially for IT grading. Internal validation was performed using a hold-out validation set and bootstrap resampling, which provided an initial estimate of model stability. However, external validation was not conducted, and the predictive performance should be interpreted as preliminary. In future research, it will be important to enlarge the dataset, apply standardized imaging protocols and automated segmentation, followed by external validation to assess clinical utility.
Conclusion
Radiologist visual assessment clearly differentiated MTs from ITs. In contrast, visual assessment alone showed limited ability to distinguish low- from high-grade ITs. Several CT- and MRI‑based radiomic features, particularly those derived from FS-T1WI, differed significantly between low- and high-grade ITs. These findings indicate that radiomics may provide objective, complementary information beyond visual assessment by capturing the signal-intensity distributions characteristic of high-grade ITs. This capability would facilitate noninvasive tumor grading and enable early treatment planning and risk stratification for recurrence and prognosis. The findings of this study will be translated into a practical tool by collecting additional cases in multicenter prospective studies, applying deep learning–based approaches, and developing a model with external validation.
Conflicts of Interest: The authors declare that they have no conflict of interest.
Supplementary Materials: Supplementary Fig. 1. Spearman correlation heatmaps of radiomic features. These Spearman correlation heatmaps illustrate the relationships among radiomics features from MRI (T2-weighted, T1-weighted, and fat-suppressed T1-weighted images) and CT. Hierarchical clustering was applied within each imaging modality to visualize feature redundancy and underlying correlation structure. 2D shape features were excluded due to their fundamentally different characteristics. Red indicates positive correlations and blue indicates negative correlations (Spearman’s ρ).
Supplementary Fig. 2. Principal component analysis of radiomic features. Scatter plots illustrate the distribution of cases in the PCA space derived from radiomics features of MRI (T2-weighted, T1-weighted, and fat-suppressed T1-weighted images) and CT. Each point represents an individual case and is color-coded by group. The first and second principal components (PC1 and PC2) are displayed, and the percentages of variance explained by each component are indicated.PCA, principal component analysis.
Supplementary Fig. 3. Mean ROC curves of LASSO and Elastic Net models for MT vs. IT discrimination and IT grading. The left panel shows MT vs. IT classification and the right panel shows low-grade vs. high-grade IT classification. For each task, mean ROC curves are shown for LASSO (cyan) and Elastic Net (pink) models, which were obtained from 200-times stratified bootstrap resampling of the training set. Shaded areas indicate the 95% confidence interval.IT, immature teratoma; LASSO, least absolute shrinkage and selection operator; MT, mature teratoma; ROC, receiver operating characteristic.
Supplementary Table 1. Group centroid shift in PCA space.
Supplementary Table 2. Selected radiomic features, feature classes, and coefficients in the final LASSO and Elastic Net models for differentiating IT from MT and high-grade from low-grade IT.
Supplementary Materials
Supplementary figures are available online.
Supplementary Table 1.
Group centroid shift in PCA space
Supplementary Table 2.
Selected radiomic features, feature classes, and coefficients in the final LASSO and Elastic Net models for differentiating IT from MT and high-grade from low-grade IT
References
- 1.Saleh M, Bhosale P, Menias CO, et al. Ovarian teratomas: clinical features, imaging findings and management. Abdom Radiol (NY) 2021; 46:2293–2307. [DOI] [PubMed] [Google Scholar]
- 2.Kaur B. Pathology of malignant ovarian germ cell tumours. Diagn Histopathol (Oxf) 2020; 26:289–297. [Google Scholar]
- 3.Moraru L, Mitranovici MI, Chiorean DM, et al. Immature teratoma: Diagnosis and management—a review of the literature. Diagnostics (Basel) 2023; 13:1516. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 4.Spinelli C, Strambi S, Masoni B, et al. Surgical management of ovarian teratomas in childhood: A multicentric study on 110 cases and a literature review. Gynecol Endocrinol 2021; 37:950–954. [DOI] [PubMed] [Google Scholar]
- 5.Marino G, Grassi T, De Ponti E, et al. Outcome of patients with stage I immature teratoma after surveillance or adjuvant chemotherapy. Front Oncol 2024; 14:1330481. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 6.Nasioudis D, Mastroyannis SA, Latif NA, Ko EM. Trends in the surgical management of malignant ovarian germcell tumors. Gynecol Oncol 2020; 157:89–93. [DOI] [PubMed] [Google Scholar]
- 7.Faure-Conter C, Pashankar F. Immature ovarian teratoma: When to give adjuvant therapy? J Pediatr Hematol Oncol 2017; 39:487–489. [DOI] [PubMed] [Google Scholar]
- 8.National Comprehensive Cancer Network. NCCN Clinical Practice Guidelines in Oncology: Ovarian Cancer Including Fallopian Tube Cancer and Primary Peritoneal Cancer (Version 3.2025). https://www.nccn.org/professionals/physician_gls/pdf/ovarian.pdf.
- 9.Pashankar F, Hale JP, Dang H, et al. Is adjuvant chemotherapy indicated in ovarian immature teratomas? A combined data analysis from the Malignant Germ Cell Tumor International Collaborative. Cancer 2016; 122:230–237. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 10.Saba L, Guerriero S, Sulcis R, Virgilio B, Melis G, Mallarini G. Mature and immature ovarian teratomas: CT, US and MR imaging characteristics. Eur J Radiol 2009; 72:454–463. [DOI] [PubMed] [Google Scholar]
- 11.Kim JC, Kim YW. CT findings of ovarian teratomas: Mature versus immature. J Korean Radiol Soc. 1996; 35:949–955. [Google Scholar]
- 12.Yamaoka T, Togashi K, Koyama T, et al. Immature teratoma of the ovary: Correlation of MR imaging and pathologic findings. Eur Radiol 2003; 13:313–319. [DOI] [PubMed] [Google Scholar]
- 13.Saida T, Mori K, Masumoto T, et al. Ovarian and non-ovarian teratomas: A wide spectrum of features. Jpn J Radiol 2021; 39:143–158. [DOI] [PubMed] [Google Scholar]
- 14.Outwater EK, Siegelman ES, Hunt JL. Ovarian teratomas: Tumor types and imaging characteristics. Radiographics 2001; 21:475–490. [DOI] [PubMed] [Google Scholar]
- 15.Zhou H, Li X, Chen J, et al. Development and validation of a nomogram for preoperative prediction of immature teratoma in children with teratoma: A retrospective, multicenter, diagnostic study. Quant Imaging Med Surg 2023; 13:8067–8078. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 16.Lambin P, Rios-Velazquez E, Leijenaar R, et al. Radiomics: Extracting more information from medical images using advanced feature analysis. Eur J Cancer 2012; 48:441–446. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 17.Nakamori A, Tsuyoshi H, Tsujikawa T, Orisaka M, Kurokawa T, Yoshida Y. Evaluation of calcification distribution by CT-based textural analysis for discrimination of immature teratoma. J Ovarian Res 2023; 16:179. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 18.van Griethuysen JJM, Fedorov A, Parmar C, et al. Computational radiomics system to decode the radiographic phenotype. Cancer Res 2017; 77:e104–e107. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 19.O’Connor DM, Norris HJ. The influence of grade on the outcome of stage I ovarian immature (malignant) teratomas and the reproducibility of grading. Int J Gynecol Pathol 1994; 13:283–289. [DOI] [PubMed] [Google Scholar]
- 20.WHO Classification of Tumours Editorial Board. Female Genital Tumours: WHO Classification of Tumours, 5th Edition, Volume 4. Lyon: International Agency for Research on Cancer, World Health Organization, 2020. [Google Scholar]
- 21.Birbas E, Kanavos T, Gkrozou F, Skentou C, Daniilidis A, Vatopoulou A. Ovarian masses in children and adolescents: A review of the literature with emphasis on the diagnostic approach. Children (Basel) 2023; 10:1114. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 22.Chen C, Li JD, Huang H, Feng YL, Wang LH, Chen L. Diagnostic value of multiple tumor marker detection for mature and immature teratoma of the ovary. Ai Zheng 2008; 27:92–95. (in Chinese). [PubMed] [Google Scholar]