Spatial habitats radiomics from multiparametric magnetic resonance imaging predict platinum resistance in epithelial ovarian cancer.

Spatial habitats radiomics from multiparametric magnetic resonance imaging predict platinum resistance in epithelial ovarian cancer - PMC Skip to main content An official website of the United States government Here's how you know Here's how you know Search Log in Primary site navigation Search Log in Search PMC Full-Text Archive Search in PMC Journal List User Guide As a library, NLM provides access to scientific literature. Inclusion in an NLM database does not imply endorsement of, or agreement with, the contents by NLM or the National Institutes of Health. Learn more: PMC Disclaimer | PMC Copyright Notice J Ovarian Res . 2026 Apr 21;19:201. doi: 10.1186/s13048-026-02097-8 Spatial habitats radiomics from multiparametric magnetic resonance imaging predict platinum resistance in epithelial ovarian cancer Lingling Lin 1 , Huawei Wu 2 , Chao Wang 3 , Jialu Xu 1 , Enhui Xin 4 , Yafen Li 5 , Jun Zhu 6 , Jianli Yu 1 , Yu Wang 3, ✉ , Jiejun Cheng 1, ✉ Author information Article notes Copyright and License information 1 Department of Radiology, Shanghai First Maternity and Infant Hospital, 2699 West Gaoke Road, Shanghai, 201204 China 2 Department of Radiology, Renji Hospital, School of Medicine, Shanghai Jiao Tong University, 1630 Dongfang Road, Shanghai, 200127 China 3 Department of Obstetrics and Gynecology, Shanghai First Maternity and Infant Hospital, 2699 West Gaoke Road, Shanghai, 201204 China 4 Department of Research and Development, Shanghai United Imaging Intelligence Co. Ltd, 701 Yunjin Road, Shanghai, 200232 China 5 University of Shanghai for Science and Technology, 1100 Jungong Road, Shanghai, 200093 China 6 National Engineering Research Center for Nanotechnology, 28 Jiangchuan East Road, Shanghai, 200241 China ✉ Corresponding author. Received 2025 Apr 22; Accepted 2026 Apr 3; Collection date 2026. © The Author(s) 2026 Open Access This article is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International License, which permits any non-commercial use, sharing, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if you modified the licensed material. You do not have permission under this licence to share adapted material derived from this article or parts of it. The images or other third party material in this article are included in the article’s Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article’s Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by-nc-nd/4.0/ . PMC Copyright notice PMCID: PMC13235077  PMID: 42015267 Abstract Objectives To investigate the potential of employing spatial habitat radiomics as an innovative tumor biomarker for predicting the response of epithelial ovarian cancer (EOC) patients to platinum-based chemotherapy. Methods and Materials In this multicentre retrospective study, 273 EOCs with preoperative MRI data (training set = 175, Hospital A; external testing set = 98, Hospital B) were included. Multivariate logistic regression analysis was conducted to identify independent predictors for the development of the clinical model. Spatial habitats were delineated using the K-means clustering method. On T2-weighted imaging and contrast-enhanced T1-weighted imaging data, 3,668 radiomics features were extracted from the whole primary tumor and each habitat. All models were constructed utilizing logistic regression, support vector machine and random forest. The area under curve (AUC), calibration plot, and decision curve were employed to assess the performance and clinical utility of models. Results Platinum resistance was detected in 48 out of 273 patients (17.6%). Multivariate regression analysis revealed that age is an independent predictor of platinum resistance in EOC patients. The primary tumors were clustered into three habitats, with Habitat 3 model showing the best predictive performance (AUC = 0.760) compared to Habitat 1 and 2 models in the testing set (both p < 0.05). The Habitat model achieved a greater AUC than the Radiomics and Clinical models (0.813 vs. 0.609 and 0.763, respectively). The habitat-based Nomogram combining 18 habitat radiomic features with clinical features had an AUC of 0.818 in the testing set, accompanied by improved calibration and clinical utility. Furthermore, the plasma lipid profile (LDL, HDL, CHOL and TG) had significant correlations with the habitat radiomics features ( p < 0.05). Conclusions The habitat-based Nomogram could predict platinum resistance in EOC patients, with habitat 3 emerging as the most predictive habitat. The habitat radiomics features may serve as indicators of plasma lipid. Supplementary Information The online version contains supplementary material available at 10.1186/s13048-026-02097-8. Keywords: Epithelial ovarian cancer, Platinum resistance, Magnetic resonance imaging, Spatial habitat radiomics, Nomogram Introduction Ovarian cancer is currently the malignant tumor with the highest mortality rate in the female reproductive system [ 1 ]. Approximately 70% of patients experience recurrence after receiving platinum-based chemotherapy, which is the first-line chemotherapy option for epithelial ovarian cancer (EOC) [ 2 , 3 ]. One of the major contributors to recurrence was the development of resistance to platinum drugs. The prognosis for EOC patients with platinum-resistant was notably poorer than that of platinum-sensitive patients, necessitating distinct treatment strategies [ 4 ]. Thus, it is crucial to accurately evaluate the sensitivity of EOC patients to platinum chemotherapy before clinical treatment. Ovarian cancer is a kind of solid tumor that exhibits genomic, immunologic, and physiological heterogeneity [ 5 – 8 ]. The treatment response varies within different regions of the tumor, and platinum resistance could develop in distinct tumor subregions [ 8 ]. An innovative method entails dividing tumors into subregions, termed habitats, which facilitates a more precise measurement of intratumoral heterogeneity [ 9 – 12 ]. The habitats clustered voxel-wise values, creating tumor subregions that may provide incremental value in the evaluation of platinum response [ 13 ]. The use of complex signatures derived from multidimensional data and radiomics analysis forms a common framework for spatial habitats radiomics [ 14 ]. The quantification of spatial habitat in terms of intratumoral heterogeneity was used to evaluate the treatment effect on breast cancer, pharyngeal cancer, and rectal cancer, yielding favorable outcomes [ 15 – 17 ]. The intratumoral heterogeneity of ovarian cancer is greater than that of breast cancer, largely due to the presence of extensive cystic components within ovarian cancer, including mucus, blood, fat, and other substances. This may be attributed to the inherent secretory function of the ovaries. Therefore, habitats may be more suitable for ovarian cancer [ 5 , 7 ]. Although apparent diffusion coefficient related histogram analysis for primary tumors is relatively straightforward, its ability to predict responses to platinum-based chemotherapy in serous ovarian cancer patients has been moderate [ 18 ]. Although classic radiomics has been employed to assess platinum-based chemotherapy efficacy for ovarian cancer, neglecting regional phenotype differences within the tumor may result in missing important information [ 19 , 20 ]. Thus, the exploration of spatial habitats’ potential in predicting platinum resistance holds value and warrants further investigation. Research has suggested that lipid metabolism plays an important role in metastasis and platinum resistance of EOC [ 21 , 22 ]. Platinum-resistant patients had higher levels of plasma low-density lipoprotein (LDL) and its receptor than platinum-sensitive patients [ 23 ]. Lipid imbalance in ovarian cancer patients may influence their sensitivity to platinum-based chemotherapy by disrupting glucose and oxygen metabolism and inflammatory reactions, leading to differences in imaging, especially in the signal, shape, and texture of magnetic resonance imaging (MRI) [ 24 – 26 ]. Studies on the relationship between high-dimensional MRI features of ovarian cancer and lipid metabolism are currently relatively sparse. Thus, exploring the correlation between tumor subregion imaging and lipid metabolism is a feasible method to validate the association between imaging and biological characteristics. Therefore, we aim to develop an MRI-based habitat radiomics model for non-invasively predicting platinum resistance and to further explore the correlation between habitat radiomics features and plasma lipids. Methods and materials Patient population The study was performed after approval from the institution’s medical research committee and the requirement for informed consent was waived (LS2308). We initially identified 1820 women with indeterminate adnexal masses via ultrasound examination at Hospital A (between January 2018 and June 2022) and Hospital B (between January 2022 and March 2023). The inclusion criteria were as follows: (1) patients underwent enhanced pelvic MRI scanning; (2) patients had surgically and pathologically confirmed EOC; (3) patients received platinum-based chemotherapy. Then 343 EOC patients who received platinum-based chemotherapy with preoperative MRI were included. The exclusion criteria were as follows: (1) loss to follow-up ( n = 32); (2) failure to complete chemotherapy ( n = 17); (3) patients who received chemotherapy or radiotherapy prior to MRI scanning ( n = 9); (4) poor MRI image quality ( n = 8); (5) incomplete clinical data ( n = 4). The training set consisted of 175 EOCs from Hospital A, and the independent testing set included 98 EOCs from Hospital B. The inclusion and exclusion criteria are presented in Fig. 1 A. Fig. 1. Open in a new tab The flowchart of the study population ( A ). The workflow of the study ( B ) Clinical endpoints and clinicopathological information All clinical endpoints for the patients were evaluated by a gynecologic oncologist (C.W., who has 10 years of experience) according to the National Comprehensive Cancer Network Guidelines [ 2 ]. All patients received standard full-course platinum-based chemotherapy after cytoreductive surgery. Persistent or progressive disease during treatment, or recurrence occurring within 6 months following the last cycle of chemotherapy, was defined as platinum resistance; Otherwise, it was defined as platinum sensitivity. The clinical variables included age, menopausal status, and CA125 level which were measured two weeks before surgery. The pathological variables included pathological subtype, surgical outcome, International Federation of Gynecology and Obstetrics (FIGO) stage, parity, ascites, peritoneal implantation, KI-67, and P53 mutation status. The overall workflow of this study As shown in Fig. 1 B, the overall workflow comprises several sequential steps: image acquisition, preprocessing, and segmentation, habitat generation, feature extraction and selection, model construction and evaluation, model comparison, and explore the association between habitat radiomics features and plasma lipid profile. The METRICS tool was systematically applied to assess the methodological rigor of the overall workflow ( https://metricsscore.github.io/metrics/METRICS.html ). The comprehensive evaluation results are summarized in Table S1 , with a score of 89.7%. Image acquisition Magnetic resonance (MR) scanning was performed using a 3.0 Tesla system (GE and Siemens Medical Systems) at Hospital A, and a 1.5 Tesla system (GE Medical Systems) at Hospital B. The MR imaging protocol mainly included sagittal T2-weighted imaging (T2WI), axial T2WI, axial fat-suppressed T2WI (FS-T2WI), axial T1-weighted imaging (T1WI), axial contrast-enhanced TIWI (CE-TIWI) and axial diffusion-weighted imaging (b = 800/1000 s/mm 2 ). This study focused on axial FS-T2WI and CE-TIWI, and the scanning details are shown in Table S2. Data preprocessing and image segmentation MR image preprocessing involved several critical steps: applying N4 bias field correction to mitigate magnetic field non-uniformity across different scanners, resampling isotropic voxels to a size of 1 mm × 1 mm × 1 mm using B-spline interpolation to address voxel inconsistencies among various sequences, and performing signal intensity normalization to minimize variations in signal strength across images obtained from different machines. The volumes of interest (VOIs), specifically delineating the primary tumor, were contoured layer by layer on the axial FS-T2WI images using ITK-SNAP (version 3.8.0) by the radiologist L.L. (with 5 years of experience). Another radiologist, J.J. (with 20 years of experience), verified all VOIs and independently delineated the primary tumors of 50 patients in the training set to assess the consistency of segmentation. Subsequently, the FS-T2WI images were registered to the corresponding CE-T1WI images, and the manually annotated VOIs from the FS-T2WI were mapped to the CE-T1WI. Habitat generation Initially, local features were analyzed within the VOI for each voxel, using CE-T1WI and FS-T2WI images respectively. A 3 × 3 × 3 moving window technique was applied to compute a 19-dimensional feature vector for every voxel. The 19 features were primarily categorized into five types, as detailed in Table S3. To determine the optimal number of clusters, a range from 3 to 10 cluster numbers was explored. The optimal number selection was based on the Calinski–Harabasz (CH) score, a metric determining clustering quality [ 27 ]. Subsequently, the K-means clustering method, applied at the population level, was employed to cluster subregions based on the local features profiles for each sample [ 14 ]. The K-means algorithm was implemented with the following parameters: initialization was performed using the “k-means++” method with 10 different centroid seeds; convergence was achieved when the relative change in the within-cluster sum of squared distances was less than 1e-4 or a maximum of 300 iterations was reached; the random state was set to “None”; and the algorithm was set to “auto”. Due to the nature of the unsupervised clustering algorithm, the K-nearest neighbors method was used to refine the characterization of voxel distribution across clusters, thereby ensuring consistency in habitat region definitions. The K-means algorithm functions by partitioning data into K distinct clusters. It iteratively updates the centroids of these clusters to minimize the sum of squares within each cluster. The central component of the K-means algorithm is the objective function, which is optimized to achieve effective clustering. This is the standard mechanism underlying the K-means algorithm, which was implemented with the parameters detailed above. is the objective function. is the number of data points. is the number of clusters. is a binary indicator (1 if data point is in cluster , 0 otherwise). is the ith data point. is the centroid of cluster . is the squared Euclidean distance between data point and centroid . Feature extraction We extracted 1834 radiomics features from the primary tumor on CE-T1WI and T2WI, respectively. For each of the three tumor habitats, 1834 features were also extracted per subregion, resulting in a total of 1834 × 3 features per modality. These features were categorized into seven types: Shape, Histogram, NGTDM, GLSZM, GLRLM, GLCM features, and Wavelet-processed features, NGTDM features, GLSZM features, GLRLM features, and GLCM features, as detailed in Table S4 . Feature extraction was performed with the pyradiomics tool, adhering to the Imaging Biomarker Standardization Initiative guidelines. Feature selection For feature selection, we calculated the intraclass correlation coefficient (ICC) in 50 patients from the training set, with VOIs independently delineated by two radiologists. Features with an ICC greater than 0.80 were retained for reliability. Statistical significance was evaluated using t-tests ( p < 0.05), and highly correlated features (Pearson’s coefficient > 0.9) were removed. Next, the minimum redundancy maximum relevance (mRMR) method was applied to reduce the number of features to 32 by balancing relevance and redundancy. Finally, the least absolute shrinkage and selection operator (LASSO) regression was employed to refine the feature selection. The optimal regularization parameter λ was determined through five-fold cross-validation based on the lowest mean square error (MSE), with the hyperparameter search conducted over a logarithmic scale ranging from 0.001 to 1 (50 evenly spaced candidate values on the log scale). Model construction and evaluation This study primarily developed the following models: the Clinical model, the traditional Radiomics model, individual habitat models (namely Habitat 1, Habitat 2, and Habitat 3), the Habitat model, and the habitat-based Nomogram. The Clinical model was constructed using independent predictor factors identified through univariate and multivariate logistic regression analyses. The traditional Radiomics model was built using radiomics features extracted from the entire primary tumor. Each individual habitat model was constructed using features extracted from the corresponding habitat subregions. The Habitat model was developed based on features from all habitat subregions. The habitat-based Nomogram was built using both the features from all habitat subregions and the clinical features. Three machine learning algorithms were selected for model development, representing three distinct and fundamental paradigms in machine learning: logistic regression (LR, a simple linear model), support vector machine (SVM, a powerful kernel-based model), and random forest (RF, a robust ensemble method). For hyperparameter optimization: LR was optimized using Bayesian search over c (inverse regularization strength) on a log scale from 1e-3 to 1e3, and penalty (l1, l2). SVM was optimized using grid search over c (1e-2 to 1e2), gamma (scale, auto), and kernel (linear, rbf). RF was optimized using random search over n_estimators (50 to 500), max_depth (3 to 15), and min_samples_split (2 to 10). The performance of different models was primarily evaluated by the area under the curve (AUC), with 95% confidence interval (CI). Additionally, other metrics including true positive (TP), true negative (TN), false positive (FP), false negative (FN), accuracy (AC), sensitivity (SE), specificity (SP), positive predictive value (PPV), negative predictive value (NPV), and F1-Score were computed [ 27 ]. Calibration curves were plotted to assess calibration accuracy. Recalibration performance was evaluated using Platt scaling, and the alignment between predicted probabilities and observed outcomes was quantified by Brier scores and expected calibration errors (ECE). Decision curve analysis (DCA) gauged the clinical utility of the predictive models, with thresholds set to identify the optimal clinical benefit. Model comparison This study conducted comparative analyses among models, specifically the Clinical model, the traditional Radiomics model, the Habitat model, and the habitat-based Nomogram, performing pairwise comparisons among these four models. Furthermore, pairwise comparisons were carried out among the three individual habitat models: Habitat 1, Habitat 2, and Habitat 3. The AUC values were compared using DeLong’s test, with the Benjamini-Hochberg correction applied for comparisons. The Integrated Discrimination Improvement (IDI) was utilized to measure the overall improvement in the prediction ability between models. Association of habitat features with plasma lipid profile The plasma lipid profiles of all patients in the testing set were collected one week before treatment. The plasma lipids included LDL, high-density lipoprotein (HDL), cholesterol (CHOL), and triglyceride (TG), which were detected by electrophoresis in the laboratory department of Hospital B. The Spearman correlation test was used to assess the correlation between habitat radiomics features and plasma lipids in the testing set. Statistical analysis For continuous variables with a normal distribution, the means and standard deviations were provided. For non-normally distributed data, medians and ranges were reported. To compare the difference of continuous variables between two groups, the independent samples t-test or Mann-Whitney U test was used as appropriate. For categorical variables, frequency (percentage) was provided. Pearson’s chi-squared tests were used to test the difference of categorical variables between two groups. The Spearman correlation test was used to assess the correlation between two variables. A larger correlation coefficient indicates a stronger relationship between the two variables, with values ranging from 0 to 1. All the statistical tests were two-sided, and the significance level was set at a P -value < 0.05. All the above analyses were performed using Python (Python 3.7, https://www.python.org/ ). Results Clinicopathological features analysis and clinical model construction The clinical and pathological characteristics of the patients in the training and testing sets are detailed in Table 1 . 36 patients (36/175, 20.6%) in the training set and 12 patients (12/98, 12.2%) in the testing set were confirmed to experience platinum resistance. Moreover, platinum resistance, age, CA125 level, menopausal status, surgical outcome, parity, ascites, KI-67, and P53 mutation were not significantly different between the training and testing sets. The FIGO stage, subtype, and peritoneal implantation were significantly different between the training and testing sets ( all P -values < 0.05). Table 1. The clinical and pathological characteristics of the patients Features ALL Training set Testing set P -value Platinum resistance 0.083 No 225(82.4) 139(79.4) 86(87.8) Yes 48(17.6) 36(20.6) 12(12.2) Age 56.44 ± 11.61 56.70 ± 11.75 55.98 ± 11.40 0.482 CA125 1024.19 ± 2858.45 1214.80 ± 3455.53 683.81 ± 1144.44 0.153 Menopausal 0.319 No 95(34.8) 56(32.0) 39(39.8) Yes 178(65.2) 119(68.0) 59(60.2) Surgical outcome 0.122 R0 228(83.5) 150(85.7) 78(79.6) R1 45(16.5) 25(14.3) 20(20.4) FIGO stage 0.002 I 64(23.4) 32(18.3) 32(32.7) II 55(20.2) 35(20.0) 20(20.4) III 108(39.6) 68(38.9) 40(40.8) IV 46(16.9) 40(22.9) 6(6.1) Subtype 0.046 HGSC 163(59.7) 102(58.3) 61(62.2) LGSC 2(0.7) None 2(2.0) OCCC 45(16.5) 26(14.9) 19(19.4) OEC 39(14.3) 29(16.6) 10(10.2) OMC 14(5.1) 8(4.6) 6(6.1) Other 10(3.7) 10(5.7) None Parity 0.987 Nultipara 162(59.3) 102(58.3) 60(61.2) Mullipara 111(40.7) 73(41.7) 38(38.8) Ascites 0.166 No 108(39.6) 64(36.6) 44(44.9) Yes 165(60.4) 111(63.4) 54(55.1) Peritoneal implantation 0.005 No 112(41.0) 59(33.7) 53(54.1) Yes 161(59.0) 116(66.3) 45(45.9) KI-67 0.51 ± 0.21 0.50 ± 0.21 0.52 ± 0.21 0.355 P53 mutation 0.441 No 81(29.7) 48(27.4) 33(33.7) Yes 192(70.3) 127(72.6) 65(66.3) Open in a new tab Categorical variables are shown as frequency (percentage), while continuous variables are shown as mean ± standard deviation; CA125 means cancer antigen 125; R0 means no residual lesions after surgery; R1 means have residual lesions after surgery; FIGO stage means International federation of gynecology and obstetrics stage; HGSC means high-grade serous ovarian cancer; LGSC means low-grade serous ovarian cancer; OCCC means ovarian clear cell carcinoma; OEC means ovarian endometrioid carcinoma; OMC means ovarian mucinous carcinoma; Other means the other subtypes of epithelial ovarian cancer Univariate logistic regression analysis revealed the following significant factors for identifying platinum resistance: age, menopausal status, FIGO stage, subtype, parity, ascites, peritoneal implantation, KI-67, and P53 mutation (Table 2 ). Multivariate regression analysis revealed that age is the only independent predictor of platinum resistance, with an OR = 0.953 (0.924–0.982) and a P- value < 0.05 (Table 2 ). Consequently, age was selected as a clinical factor in subsequent analysis. The performance of all the Clinical models was presented in Table S5. Among the three machine learning models, the clinical model based on LR had the best AUC of 0.763 in the testing set (Fig. S1 ). Table 2. Univariable and multivariable analyses of clinical features Features OR(95%CI) P -value OR(95%CI) P -value KI-67 0.095(0.053–0.173) < 0.001 0.261(0.055–1.232) 0.154 Menopausal 0.253(0.173–0.368) < 0.001 1.298(0.491–3.428) 0.659 P53 mutation 0.309(0.219–0.436) < 0.001 1.071(0.499–2.298) 0.883 Peritoneal implantation 0.381(0.271–0.536) < 0.001 1.385(0.402–4.778) 0.665 Ascites 0.405(0.287–0.572) < 0.001 3.154(1.188–8.381) 0.053 Parity 0.470(0.383–0.576) < 0.001 0.823(0.41–1.649) 0.645 Subtype 0.636(0.540–0.750) < 0.001 1.082(0.869–1.649) 0.552 FIGO stage 0.650(0.561–0.752) < 0.001 1.727(1.026–2.907) 0.084 Age 0.977(0.972–0.983) < 0.001 0.953(0.924–0.982) < 0.001 Surgical outcome 0.667(0.341–1.305) 0.321 NA NA CA125 1.000(1.000–1.000) 0.153 NA NA Open in a new tab OR means odds ratio; CI means confidence interval; FIGO stage means International Federation of Gynecology and Obstetrics stage; CA125 means cancer antigen 125 Radiomics feature analysis and Radiomics model construction The details of the radiomic feature analysis and radiomic model construction are shown in Fig. S2. The LASSO regularization parameter λ was optimized to 0.0256 through cross-validation based on the MSE (Fig. S2A, B). The performance of all the Radiomic models was presented in Table S5. Among the three machine learning algorithms, the radiomics model based on RF had the best AUC of 0.609 in the testing set (Fig. S2C), whereas its AUC was 0.713 in the training set (Fig. S2D). The coefficients of 19 selected radiomics features in the model were listed in sequence (Fig. S2E). Spatial habitat features analysis and Habitat model construction The optimal number of clusters was determined to be three based on the CH score (Fig. S3A). The visualization of the clustering results is shown in (Fig. S3B), where the distinct separation of points in different colors indicates a favorable outcome for clustering. Thus, the primary tumor of EOC patient was clustered into three habitats, illustrated in Fig. 2 A. The habitat group included features from 3 habitats, totaling 11,004 features (1,834 * 3 * 2), categorized into seven groups (Fig. S3C). The LASSO regularization parameter λ was optimized to 0.0193 through cross-validation based on the MSE (Fig. S3D). The Habitat model based on RF had the best AUC of 0.813 in the testing set (Fig. S3F), whereas its AUC was 0.944 in the training set (Fig. S3E). The coefficients of 18 selected radiomics features in the model were listed in sequence (Fig. S3G). The performance of all the Habitat models was presented in Table S5. Fig. 2. Open in a new tab The case of EOC patient demonstrating the spatial habitat subregions of the whole primary tumor for predicting platinum resistance is presented ( A ). The voxel values and volume for each habitat are shown ( B ). The first and second columns show the distribution of voxel values for each habitat on T2WI and CE-T1WI, as well as their distribution in the PR and PS groups; The third column shows the distribution of voxel volume for each habitat, as well as their distribution in the PR and PS groups. The AUCs of Habitat 1, 2 and 3 models in the training ( C ) and testing sets ( D ) are shown The details of the habitats within the tumor are as follows. Figure 2 B presents the voxel value of each habitat in T2WI and CE-T1WI, and voxel volume of each habitat. The voxel values for habitat 3 in platinum-resistant patients were significantly lower than those in the platinum-sensitive group for both T2WI and CE-T1WI (both P -values < 0.05). The voxel volume of each habitat had no significant difference. Figure 2 C, D show the AUCs of each individual habitat model in the training and testing sets. In the testing set, the Habitat 3 model achieved the highest AUC of 0.760, significantly outperforming the Habitat 1 model (AUC = 0.582) and the Habitat 2 model (AUC = 0.661), with all pairwise comparisons being statistically significant ( all P -values < 0.05). Model comparison and habitat-based Nomogram construction The predictive performance (AUC, TP, TN, FP, FN, AC, SE, SP, PPV, NPV, F1 Score) of the Clinical model, the traditional Radiomics model, the Habitat model, and the habitat-based Nomogram are summarized in Table 3 . The Clinical model exhibited AUC values of 0.762 and 0.763 in the training and testing sets, respectively. The Radiomics model achieved AUC values of 0.713 and 0.609 in the training and testing sets, respectively. In comparison, the Habitat model yielded AUC values of 0.944 (95% CI: [0.912–0.976]) and 0.813 (95% CI: [0.735–0.891]) in the training and testing sets, respectively. Furthermore, the habitat-based Nomogram achieved an AUC of 0.944 (95% CI: [0.912–0.976]) in the training set and 0.818 (95% CI: [0.714–0.922]) in the testing set. The AUCs of the Clinical, Radiomics, Habitat, and Nomogram models in the training and testing sets are presented in Fig. 3 A, B, with the habitat-based Nomogram exhibiting the highest AUC. Stratified analyses of the model’s performance, based on FIGO stage, peritoneal implantation, and tumor pathological subtype, showed that the AUC values of the models ranged from 0.746 to 1.0 across all subgroups. Table 3. Performance of the optimal models across different strategies for predicting platinum resistance Model Data Set AUC (95% CI) TP TN FP FN AC SE SP PPV NPV F1_ Score Clinical Train 0.762 (0.681–0.843) 21 103 36 15 0.709 0.583 0.741 0.368 0.873 0.452 Clinical Test 0.763 (0.613–0.912) 8 60 26 4 0.694 0.667 0.698 0.235 0.938 0.348 Radiomics Train 0.713 (0.613–0.814) 21 109 30 15 0.743 0.583 0.784 0.412 0.879 0.483 Radiomics Test 0.609 (0.446–0.771) 8 41 45 4 0.500 0.667 0.477 0.151 0.911 0.246 Habitat Train 0.944 (0.913–0.976) 34 111 28 2 0.829 0.944 0.799 0.548 0.982 0.694 Habitat Test 0.813 (0.704–0.922) 8 61 25 4 0.704 0.667 0.709 0.242 0.938 0.356 Nomogram Train 0.944( 0.912–0.976) 36 75 64 0 0.634 1.000 0.540 0.360 1.000 0.529 Nomogram Test 0.818 (0.714–0.922) 10 59 27 2 0.704 0.833 0.686 0.270 0.967 0.408 Open in a new tab Note: AUC means the area under the receiver operating characteristic curve; CI means confidence interval; TP means true positive; TN means true negative; FP means false positive; FN means false negative; AC means accuracy; SE means sensitivity; SP means specificity; PPV means positive predictive value; NPV means negative predictive value Fig. 3. Open in a new tab The analyses of each model performance. The AUCs of the Clinical, Radiomics, Habitat model and habitat-based Nomogram in training and testing sets ( A , B ); The calibration plot of the Clinical, Radiomics, Habitat model and habitat-based Nomogram in training and testing sets ( C , D ); The decision curve analysis (DCA) of habitat-based Nomogram in training and testing sets, with a threshold of 0.105 ( E , F ); The composition of habitat-based Nomogram is shown in ( G ) Calibration curves for all models are shown in Fig. 3 C, D: the Habitat model and Nomogram curves were closest to the perfect line in the training set, while the Nomogram curve was closest in the testing set. Recalibration curves are provided in Fig. S4A, B. Table S6 summarizes original and calibrated Brier scores and ECE. Post-recalibration, the Radiomics and Habitat models showed substantial calibration improvements; the Nomogram improved in the training set but varied slightly in the testing set; and the Clinical model performance declined. The DCA of the habitat-based Nomogram exhibited good net clinical benefit in the training and testing sets (Fig. 3 E, F). Specifically, the DCA confirmed that the predefined high risk threshold (predicted probability ≥ 0.105) yields optimal net benefit within the clinically relevant range, supporting the utility of this binary classification rule. The details of the habitat-based Nomogram are shown in Fig. 3 G. The Delong analyses showed that the Nomogram had a better AUC than the Clinical and Radiomic models in the training set ( all P -values < 0.05) but the three models had no significant difference in the testing set. (Fig. S4C, D). The comparison of the IDIs revealed the enhanced prediction ability of the Nomogram over the Clinical model and Radiomics model (Fig. S4E, F). Associations between spatial habitats and plasma lipid profile The plasma lipid levels (LDL, HDL, CHOL, and TG) of the patients in the testing set are detailed in Table S7. There were significant differences in LDL between the platinum-resistant group and the platinum-sensitive group ( p < 0.05). These two groups had no significant differences in HDL, CHOL, or TG. The correlations between habitat radiomics features and plasma lipids (LDL, HDL, CHOL, and TG) were analyzed (Fig. 4 ). LDL was significantly correlated with the wavelet-HHL_glcm_ClusterShade_h3_T1WI. HDL was significantly correlated with two habitat radiomics features. CHOL was significantly correlated with two habitat radiomics features. TG had significant correlations with six habitat radiomics features. Fig. 4. Open in a new tab The correlation coefficients between habitat characteristics and plasma lipid (LDL, HDL, CHOL, and TG) are shown in the figure, and “*” means the correlation between the two factors was significant Discussion We constructed a habitat nomogram to predict platinum resistance in EOC patients. The habitat-based Nomogram, which integrates MRI habitat radiomics features and clinicopathological characteristics, achieved an AUC of 0.818 in the external testing set, with favorable calibration ability and clinical utility. Notably, Habitat 3 model exhibited the strongest association with platinum resistance, achieving the highest AUC of 0.760 in the testing set and highlighting habitat 3’s potential as a key imaging subregion for predicting platinum response. Additionally, it is the first time that habitat radiomics features have been demonstrated to be associated with plasma lipids, revealing the relationships between habitat images and lipid-related biomarkers. Unlike previous radiomic studies on entire tumors, an emerging approach explicitly aimed at identifying distinct tumor subregions or cell subpopulations is commonly called habitat imaging, which has been reported to reveal aggressive subregions that are important for determining prognosis and treatment response [ 10 , 27 ]. The subregion cluster analysis performed in this study demonstrated the importance of subregion analysis in reflecting the tumor heterogeneity of EOC patients, which was better than classical radiomic analysis. This result is consistent with previous studies that suggested the prognostic role of subregional heterogeneity [ 14 ]. K-means, Otsu, and Gaussian Mixture Model (GMM) clustering are widely used methods for MRI data clustering [ 28 – 30 ]. However, the Otsu method may not fully capture the complexities of tumor pathophysiology, including aspects such as cell density and spatial heterogeneity [ 29 ]. On the other hand, while GMM clustering offers flexibility, it is computationally intensive and often struggles with high-dimensional data [ 30 ]. In this study, we employed K-means clustering, which iteratively groups data points based on the mean of centroids. This method is straightforward, easy to reproduce, computationally efficient, and highly interpretable, making it a practical choice for our analysis [ 28 ]. Habitat imaging postulated that subregions identified by voxels with similar imaging traits share common tumor biology [ 31 , 32 ]. The voxels from the solid and cystic areas could be different. Based on the signal intensities of each habitat subregion on CE-T1WI and T2WI as shown in Fig. 2 , we speculate that: tumors are mainly divided into three habitats representing different biological characteristics. Specifically, the red habitat (habitat 1) mainly corresponds to the dense area of the solid component, which contains an abundance of blood vessels (high CE-T1WI and low T2WI signal intensity), the green habitat (habitat 2) mainly corresponds to the cystic area, and the blue habitat (habitat 3) mainly corresponds to the loose area of the solid component, which contains sparse blood vessels (low CE-T1WI and high T2WI signal intensity). Furthermore, the habitat 3 exhibited a stronger association with platinum resistance in EOC compared to other habitats. This conclusion was supported by the following evidence: First, the Habitat 3 model achieved the highest AUC among the three models evaluated. Second, the habitat-based Nomogram incorporated 11 features (11/18, 61%) extracted specifically from the habitat 3. Third, the habitat radiomic feature square_glcm_cluster shade_h3_T1WI, derived from the habitat 3, demonstrated the highest relative contribution among all features utilized in the Habitat model. Whether on T2WI or CE-T1WI, the voxel values of the habitat 3 show significant differences between platinum-sensitive and platinum-resistant patients. The primary tumor sites of platinum-resistant patients exhibit low signal intensity and weak enhancement degree, with hypoxic regions characterized by sparse vasculature, making them more prone to developing resistance to platinum-based chemotherapy [ 33 ]. Therefore, these findings suggested that the loose area of the solid component is more strongly associated with platinum resistance in EOC compared to other areas. This aligned with a study identifying the loose area of the solid component and/or the border zone between the solid and cystic components as the most critical area in the tumor for differential diagnosis of ovarian cancer subtypes [ 34 ]. Identifying the critical subregion reflecting platinum resistance in ovarian cancer could facilitate the selection of target cancer cells for drug sensitivity testing, presenting a potential clinical application for the habitat nomogram. Although the habitat-based nomogram demonstrated significant incremental value over the clinical model in the training set, this superiority was not statistically confirmed in the testing set. This outcome may be attributed to the limited sample size of the testing set or cohort heterogeneity. Stratified analyses were conducted to explore the impact of cohort heterogeneity on the nomogram. The AUCs had a relatively wide range across all subgroups, ranging from 0.746 to 1. Further investigation showed that the AUCs reached 1 in both the early FIGO stage subgroup and the non-peritoneal implantation subgroup; this was likely because only one patient in each of these subgroups was platinum-resistant, and substantial variation in platinum resistance rates led to an overestimated AUC. Excluding these two subgroups with low platinum resistance rates, the AUCs of the remaining subgroups ranged from 0.746 to 0.875, which was comparable to the nomogram’s AUC in the testing set (0.818). Besides, multivariable logistic regression analysis showed the nomogram predicted probability emerged as the sole independent predictor (odds ratio (OR) = 4900.80, 95% CI [23.78–1,010,166.10], p = 0.002), fully encompassing key clinical variables. These suggest that the nomogram was less affected by cohort heterogeneity and may be applicable to diverse clinical scenarios. Nine clinical variables were significantly associated with platinum resistance in univariate regression, but only age remained an independent predictor in multivariate regression. Consistent with a previous study [ 19 ], the OR of age was close to 1, indicating its limited predictive ability for platinum resistance. These results indicate the limitations of using only clinical features to predict platinum resistance. Integrating clinical and imaging features into a multi-dimensional model may better assess ovarian cancer patients, with larger multi-center validation needed. Our study found significantly elevated LDL levels in platinum-resistant ovarian cancer, confirming the established link between lipid metabolism and chemoresistance [ 23 ]. Critically, we identified the association between LDL and the habitat radiomics feature Wavelet_HLL_glcm_ClusterShade (from primary tumor habitat 3), which enabled a biologically grounded interpretation of the imaging marker. Wavelet_HLL_glcm_ClusterShade is a measure of the skewness and uniformity of the Gray-Level Co-occurrence Matrix [ 35 ], reflecting high textural heterogeneity on MRI. We hypothesize this heterogeneity mirrors lipid metabolism-driven biological dysregulation: elevated values may correspond to lipid-rich necrotic or apoptotic cell clusters interspersed with viable tumor cells, and disorganized tissue architecture (lipid-filled cysts, dense cellularity) that impairs platinum delivery. Therefore, systemic lipid dysregulation is locally mirrored as a heterogeneous tumor habitat, non-invasively detectable via MRI. This reinforces that habitat imaging reflects the functional tumor microenvironment, and our model’s prognostic power likely stems from quantifying this lipid-associated high-heterogeneity phenotype. Future spatial transcriptomics studies will validate these links to lipid pathway molecular alterations. Combining habitat radiomics with lipid features is also under consideration. The DCA results reinforced the clinical relevance of the Nomogram. The superior net benefit observed at the ≥ 0.105 threshold aligns with our emphasis on minimizing false negatives—a critical priority in first-line EOC treatment, where missed platinum resistance could lead to ineffective therapy and delayed optimal intervention. This finding supported the integration of the Nomogram’s binary classification (High/Low Risk) into clinical workflows, providing an evidence-based tool to balance sensitivity and specificity while prioritizing patient safety. Calibration results indicated that recalibration notably enhanced Radiomics and Habitat models, with substantial ECE reductions confirming corrected miscalibration in imaging-based outputs. For the Nomogram, training set improvements supported recalibration value, but testing set variability and strong baseline performance justified retaining original probabilities for final metrics—ensuring clinical clarity. A previous study on MRI-based habitat analysis for predicting platinum resistance has only focused on high-grade serous ovarian carcinoma, neglecting other EOC subtypes treated with first-line platinum-based chemotherapy, including endometrioid, clear cell, and mucinous ovarian carcinomas [ 36 ]. All platinum-eligible EOC subtypes were included in this study to improve the clinical utility of model. Moreover, our study unearthed significant correlations between LDL and selected habitat features of the habitat-based Nomogram, aligning with prior studies [ 27 ]. This correlation provided preliminary insights into the biological underpinnings of our nomogram. Another study predicted platinum resistance in ovarian cancer using a CT radiomic model based on pre-treatment images and genomic data, including Single-Nucleotide Polymorphisms of Human Sulfatase 1, achieving an AUC of 0.967 in the validation cohort [ 19 ]. However, the reliance on costly genetic testing complicates clinical translation. Furthermore, as both training and validation cohorts were sourced from the same institution, this limits the model’s generalizability. The limitations of this study are as follows. Firstly, our use of three-dimensional VOIs for tumor segmentation is time-consuming, which may affect the generalizability of the nomogram. We will further develop a precise automated algorithm for ovarian cancer segmentation to reduce labor costs and minimize subjective bias. Secondly, the small sample size and cohort heterogeneity may affect model performance. However, our multivariable analysis mitigated this concern: the nomogram’s predicted probability remained the sole independent predictor, fully encompassing key clinical variables. This suggested robustness despite cohort differences, though validation in larger, more balanced multicenter cohorts is still needed to confirm stability. Thirdly, while N4 bias field correction addressed magnetic field inhomogeneity, it did not resolve inter-scanner or inter-protocol variabilities. Though the model showed some robustness to this heterogeneity across institutions, such variability remains a key challenge for radiomics reproducibility. Future studies should explore harmonization methods like ComBat or standardized protocols to improve generalizability. Finally, the spatially distinct habitats identified through K-means clustering could not be pathologically confirmed due to technical challenges and ethical considerations, as pursuing such a correlation might entail unnecessary surgeries. To address these issues, we will investigate the correlation between pathology and imaging in murine models by employing three-dimensional printing to create accurate morphological representations of ovarian cancer, conducting meticulously layered annotations, and extracting pathological samples. This approach aims to elucidate the correlations between pathology and MR imaging at a sub-regional level. Conclusion In conclusion, we evaluated and quantified the spatial heterogeneity of EOC, utilizing a habitat map from preoperative MR images. The habitat-based Nomogram served as an independent predictor for platinum resistance, holding promise for advancing precision medicine and shaping treatment strategies. The correlation between habitat radiomics features and plasma lipids, not only validated the close association between imaging and biological characteristics but also provided preliminary insights into the biological underpinnings of our nomogram. Supplementary Information Supplementary Material 1. (2MB, docx) Acknowledgements The authors would like to thank Shushan Dong for her assistance with the manuscript revision. Abbreviations EOC Epithelial ovarian cancer LDL Low-density lipoprotein CA125 Cancer antigen 125 FIGO International Federation of Gynecology and Obstetrics T2WI T2-weighted imaging FS-T2WI Fat-suppressed T2WI T1WI T1-weighted imaging CE-TIWI Contrast-enhanced TIWI VOIs Volumes of interest CH Calinski–Harabasz ICC Intraclass correlation coefficient mRMR Minimum redundancy maximum relevance LASSO Least absolute shrinkage and selection operator MSE Lowest mean square error LR Logistic regression SVM Support vector machine RF Random forest AUC Area under the curve TP True positive TN True negative FP False positive FN False negative AC Accuracy SE Sensitivity SP Specificity PPV Positive predictive value NPV Negative predictive value ECE Expected calibration errors DCA Decision curve analysis IDI Integrated discrimination improvement HDL High-density lipoprotein CHOL Cholesterol TG Triglyceride Authors' contributions Conceptualization, Lingling Lin; Data curation, Huawei Wu; Formal analysis, Lingling Lin; Funding acquisition, Lingling Lin and Jiejun Cheng; Investigation, Chao Wang and Jialu Xu; Methodology, Lingling Lin and Enhui Xin; Project administration, Jiejun Cheng and Yu Wang; Resources, Huawei Wu and Jiejun Cheng; Software, Lingling Lin and Yafen Li; Supervision, Jiejun Cheng and Yu Wang; Validation, Jiejun Cheng and Yu Wang; Visualization, Jianli Yu and Jun Zhu; Writing - original draft, Lingling Lin; Writing - review & editing, Jiejun Cheng. Funding This study has received funding from clinical research special project of Shanghai First Maternity and Infant Hospital (2023B16), Shanghai Municipal Science and Technology Commission Explorer Project (23TS1401000), Tongji University Medical-Engineering Interdisciplinary Program (2025-0555-YB-01), and Talent Introduction of Shanghai First Maternity and Infant Hospital Project (2022RC04). Data availability The experimental data used to support the findings of this study are available from the Corresponding author upon request. Declarations Ethics approval and consent to participate This study was approved by the Ethics Committee of Shanghai First Maternity and Infant Hospital (LS2308). Data from Renji Hospital, School of Medicine, Shanghai Jiao Tong University were used under a data-sharing agreement authorized by the approving committee. The requirement for informed consent was waived by the ethics committee for all data, as the study involved only anonymized retrospective analysis. Competing interests The authors declare no competing interests. Footnotes Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. Contributor Information Yu Wang, Email: renjiwangyu@126.com. Jiejun Cheng, Email: yfyfskcjj@163.com. References 1. Siegel RL, Miller KD, Wagle NS, Jemal A. Cancer statistics, 2023. CA Cancer J Clin. 2023;73(1):17–48. [ DOI ] [ PubMed ] [ Google Scholar ] 2. Armstrong DK, Alvarez RD, Backes FJ, Bakkum-Gamez JN, Barroilhet L, Behbakht K, et al. NCCN Guidelines ® Insights: Ovarian Cancer, Version 3.2022. J Natl Compr Canc Netw. 2022;20(9):972–80. [ DOI ] [ PubMed ] [ Google Scholar ] 3. Richardson DL, Eskander RN, O’Malley DM. Advances in Ovarian Cancer Care and Unmet Treatment Needs for Patients With Platinum Resistance: A Narrative Review. JAMA Oncol. 2023;9(6):851–9. [ DOI ] [ PubMed ] [ Google Scholar ] 4. Fabbro M, Colombo PE, Leaha CM, Rouanet P, Carrère S, Quenet F, et al. Conditional Probability of Survival and Prognostic Factors in Long-Term Survivors of High-Grade Serous Ovarian Cancer. Cancers (Basel). 2020;12(8):2184. [ DOI ] [ PMC free article ] [ PubMed ] 5. Vázquez-García I, Uhlitz F, Ceglia N, Lim JLP, Wu M, Mohibullah N, et al. Ovarian cancer mutational processes drive site-specific immune evasion. Nature. 2022;612(7941):778–86. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 6. Zhang D, Lu W, Cui S, Mei H, Wu X, Zhuo Z. Establishment of an ovarian cancer omentum metastasis-related prognostic model by integrated analysis of scRNA-seq and bulk RNA-seq. J Ovarian Res. 2022;15(1):123. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 7. Liang L, Yu J, Li J, Li N, Liu J, Xiu L, et al. Integration of scRNA-Seq and Bulk RNA-Seq to Analyse the Heterogeneity of Ovarian Cancer Immune Cells and Establish a Molecular Risk Model. Front Oncol. 2021;11:711020. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 8. Yu L, Ding Y, Wan T, Deng T, Huang H, Liu J. Significance of CD47 and Its Association With Tumor Immune Microenvironment Heterogeneity in Ovarian Cancer. Front Immunol. 2021;12:768115. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 9. Park JE, Kim HS, Kim N, Park SY, Kim YH, Kim JH. Spatiotemporal Heterogeneity in Multiparametric Physiologic MRI Is Associated with Patient Outcomes in IDH-Wildtype Glioblastoma. Clin Cancer Res. 2021;27(1):237–45. [ DOI ] [ PubMed ] [ Google Scholar ] 10. Wu J, Cao G, Sun X, Lee J, Rubin DL, Napel S, et al. Intratumoral Spatial Heterogeneity at Perfusion MR Imaging Predicts Recurrence-free Survival in Locally Advanced Breast Cancer Treated with Neoadjuvant Chemotherapy. Radiology. 2018;288(1):26–35. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 11. Zhang Y, Yang C, Sheng R, Dai Y, Zeng M. Predicting the recurrence of hepatocellular carcinoma (≤ 5 cm) after resection surgery with promising risk factors: habitat fraction of tumor and its peritumoral micro-environment. Radiol Med. 2023;128(10):1181–91. [ DOI ] [ PubMed ] [ Google Scholar ] 12. Lee DH, Park JE, Kim N, Park SY, Kim YH, Cho YH, et al. Tumor Habitat Analysis Using Longitudinal Physiological MRI to Predict Tumor Recurrence After Stereotactic Radiosurgery for Brain Metastasis. Korean J Radiol. 2023;24(3):235–46. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 13. O’Connor JP, Rose CJ, Waterton JC, Carano RA, Parker GJ, Jackson A. Imaging intratumor heterogeneity: role in therapy response, resistance, and clinical outcome. Clin Cancer Res. 2015;21(2):249–57. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 14. Huang H, Chen H, Zheng D, Chen C, Wang Y, Xu L, et al. Habitat-based radiomics analysis for evaluating immediate response in colorectal cancer lung metastases treated by radiofrequency ablation. Cancer Imaging. 2024;24(1):44. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 15. Shi Z, Huang X, Cheng Z, Xu Z, Lin H, Liu C, et al. MRI-based Quantification of Intratumoral Heterogeneity for Predicting Treatment Response to Neoadjuvant Chemotherapy in Breast Cancer. Radiology. 2023;308(1):e222830. [ DOI ] [ PubMed ] [ Google Scholar ] 16. Wu J, Gensheimer MF, Zhang N, Guo M, Liang R, Zhang C, et al. Tumor Subregion Evolution-Based Imaging Features to Assess Early Response and Predict Prognosis in Oropharyngeal Cancer. J Nucl Med. 2020;61(3):327–36. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 17. Daye D, Tabari A, Kim H, Chang K, Kamran SC, Hong TS, et al. Quantitative tumor heterogeneity MRI profiling improves machine learning-based prognostication in patients with metastatic colon cancer. Eur Radiol. 2021;31(8):5759–67. [ DOI ] [ PubMed ] [ Google Scholar ] 18. Lu J, Li HM, Cai SQ, Zhao SH, Ma FH, Li YA, et al. Prediction of Platinum-based Chemotherapy Response in Advanced High-grade Serous Ovarian Cancer: ADC Histogram Analysis of Primary Tumors. Acad Radiol. 2021;28(3):e77–85. [ DOI ] [ PubMed ] [ Google Scholar ] 19. Yi X, Liu Y, Zhou B, Xiang W, Deng A, Fu Y, et al. Incorporating SULF1 polymorphisms in a pretreatment CT-based radiomic model for predicting platinum resistance in ovarian cancer treatment. Biomed Pharmacother. 2021;133:111013. [ DOI ] [ PubMed ] [ Google Scholar ] 20. Li H, Cai S, Deng L, Xiao Z, Guo Q, Qiang J, et al. Prediction of platinum resistance for advanced high-grade serous ovarian carcinoma using MRI-based radiomics nomogram. Eur Radiol. 2023;33(8):5298–308. [ DOI ] [ PubMed ] [ Google Scholar ] 21. Wang Y, Calvert AE, Cardenas H, Rink JS, Nahotko D, Qiang W, et al. Nanoparticle Targeting in Chemo-Resistant Ovarian Cancer Reveals Dual Axis of Therapeutic Vulnerability Involving Cholesterol Uptake and Cell Redox Balance. Adv Sci (Weinh). 2024;11(13):e2305212. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 22. Criscuolo D, Avolio R, Calice G, Laezza C, Paladino S, Navarra G, et al. Cholesterol Homeostasis Modulates Platinum Sensitivity in Human Ovarian Cancer. Cells. 2020;9(4):828. [ DOI ] [ PMC free article ] [ PubMed ] 23. Huang X, Wei X, Qiao S, Zhang X, Li R, Hu S, et al. Low Density Lipoprotein Receptor (LDLR) and 3-Hydroxy-3-Methylglutaryl Coenzyme a Reductase (HMGCR) Expression are Associated with Platinum-Resistance and Prognosis in Ovarian Carcinoma Patients. Cancer Manag Res. 2021;13:9015–24. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 24. Tan Y, Li J, Zhao G, Huang KC, Cardenas H, Wang Y, et al. Metabolic reprogramming from glycolysis to fatty acid uptake and beta-oxidation in platinum-resistant cancer cells. Nat Commun. 2022;13(1):4554. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 25. Rameh V, Vajapeyam S, Ziaei A, Kao P, London WB, Baker SJ, et al. Correlation between Multiparametric MR Imaging and Molecular Genetics in Pontine Pediatric High-Grade Glioma. AJNR Am J Neuroradiol. 2023;44(7):833–40. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 26. Buechel ME, Enserro D, Burger RA, Brady MF, Wade K, Secord AA, et al. Correlation of imaging and plasma based biomarkers to predict response to bevacizumab in epithelial ovarian cancer (EOC). Gynecol Oncol. 2021;161(2):382–8. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 27. Xie C, Yang P, Zhang X, Xu L, Wang X, Li X, et al. Sub-region based radiomics analysis for survival prediction in oesophageal tumours treated by definitive concurrent chemoradiotherapy. EBioMedicine. 2019;44:289–97. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 28. Yuan J, Wu M, Qiu L, Xu W, Fei Y, Zhu Y, et al. Tumor habitat-based MRI features assessing early response in locally advanced nasopharyngeal carcinoma. Oral Oncol. 2024;158:106980. [ DOI ] [ PubMed ] [ Google Scholar ] 29. Qi M, Zhou W, Yuan Y, Song Y, Zhang D, Ren J. Computed tomography radiomics reveals prognostic value of immunophenotyping in laryngeal squamous cell carcinoma: a comparison of whole tumor- versus habitats-based approaches. BMC Med Imaging. 2024;24(1):304. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 30. Wang X, Deng C, Kong R, Gong Z, Dai H, Song Y. al. Intratumoral and peritumoral habitat imaging based on multiparametric MRI to predict cervical stromal invasion in early-stage endometrial carcinoma. Acad Radiol 2024 Oct 4:S1076-6332(24)00689-5. [ DOI ] [ PubMed ] 31. Lee DH, Park JE, Kim N, Park SY, Kim YH, Cho YH, et al. Tumor habitat analysis by magnetic resonance imaging distinguishes tumor progression from radiation necrosis in brain metastases after stereotactic radiosurgery. Eur Radiol. 2022;32(1):497–507. [ DOI ] [ PubMed ] [ Google Scholar ] 32. Kim M, Park JE, Kim HS, Kim N, Park SY, Kim YH, et al. Spatiotemporal habitats from multiparametric physiologic MRI distinguish tumor progression from treatment-related change in post-treatment glioblastoma. Eur Radiol. 2021;31(8):6374–83. [ DOI ] [ PubMed ] [ Google Scholar ] 33. Cummings M, Freer C, Orsi NM. Targeting the tumour microenvironment in platinum-resistant ovarian cancer. Semin Cancer Biol. 2021;77:3–28. [ DOI ] [ PubMed ] [ Google Scholar ] 34. Jian J, Li Y, Pickhardt PJ, Xia W, He Z, Zhang R, et al. MR image-based radiomics to differentiate type Ι and type ΙΙ epithelial ovarian cancers. Eur Radiol. 2021;31(1):403–10. [ DOI ] [ PubMed ] [ Google Scholar ] 35. Lambin P, Rios-Velazquez E, Leijenaar R, Carvalho S, van Stiphout RG, Granton P, et al. Radiomics: extracting more information from medical images using advanced feature analysis. Eur J Cancer. 2012;48(4):441–6. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 36. Bi Q, Miao K, Xu N, Hu F, Yang J, Shi W, et al. Habitat Radiomics Based on MRI for Predicting Platinum Resistance in Patients with High-Grade Serous Ovarian Carcinoma: A Multicenter Study. Acad Radiol. 2024;31(6):2367–80. [ DOI ] [ PubMed ] [ Google Scholar ] Associated Data This section collects any data citations, data availability statements, or supplementary materials included in this article. Supplementary Materials Supplementary Material 1. (2MB, docx) Data Availability Statement The experimental data used to support the findings of this study are available from the Corresponding author upon request. Articles from Journal of Ovarian Research are provided here courtesy of BMC ACTIONS View on publisher site PDF (2.2 MB) Cite Collections Permalink RESOURCES Similar articles Cited by other articles Links to NCBI Databases Follow NCBI NCBI on X (formerly known as Twitter) NCBI on Facebook NCBI on LinkedIn NCBI on GitHub NCBI RSS feed Connect with NLM NLM on X (formerly known as Twitter) NLM on Facebook NLM on YouTube National Library of Medicine 8600 Rockville Pike Bethesda, MD 20894 Web Policies FOIA HHS Vulnerability Disclosure Help Accessibility Careers NLM NIH HHS USA.gov Back to Top

Related documents

• Radiomics-Based AI for the Diagnosis and Prognosis of Vessels Encapsulating Tumor Clusters in Hepatocellular Carcinoma: Systematic Review and Meta-Analysis.#340368
• ECLECTIC: EstroTEP and Circulating Biomarkers for ER-positive HER2-negative Metastatic Breast Cancer Patients#1114334
• Radiomics and Radiogenomics Models to Predict Molecular Integrated Risk Classes and Prognostic Factors in Endometrial Cancer.#892038
• Preliminary Study on the Diagnostic Value Comparison of 18F-FAPI-04 PET/CT and PET/MR in Gastric Cancer#842920
• Feasibility of Serial Multisite Image-guided Biopsy to Study Breast Cancer Evolution#898848