2.1 Participants and Procedures

Data were collected from records of patients with pathologically confirmed primary HGSOC in our two affiliated hospitals between 3 January 2010 and 28 December 2020. To be eligible, patients had to undergo cytoreductive surgery following ¹⁸F-FDG PET/CT scan, the surgery had to take place at our facilities, and patients had to be pathologically diagnosed with HGSOC. Criteria for exclusion included failure to identify primary ovarian lesions on PET or CT images, presence of significant artifacts in CT images hindering tumor observation, inability to detect metastasis or absence of metastatic lesions, and incomplete data. Platinum-based chemotherapy was administered to all the patients after surgery. Platinum resistance was characterized by a period of < 6 months without platinum treatment following the initial therapy [27].

2.2 Image Acquisition

All participants underwent PET/CT imaging utilizing the Discovery PET/CT 690 system (GE Healthcare, USA) following established guidelines [28, 29]. The ¹⁸F-FDG utilized in the scans was manufactured by the hospital, boasting a purity exceeding 99% and a pH range of 4.5–8.5. Prior to the PET/CT examination, participants were required to fast for a minimum of 6 h and have their blood glucose levels assessed. Blood glucose levels were required to be below 7 mmol/L, and the injection of ¹⁸F-FDG (3.7 MBq/kg) was administered through the brachial vein. PET scans should be conducted at a rate of 1.5 min per bed, with a total of 8–9 beds acquired, encompassing the anatomical range from the apex of the skull to the mid-femur. Spiral CT scanning should utilize a layer thickness of 3.8 mm, a tube voltage of 120 kV, and an automatic tube current ranging from 15 to 180 mA. Reconstruction of attenuation-corrected PET images should employ the ordered subset expectation maximization (OSEM) algorithm with 24 subsets and two iterations, in addition to utilizing time-of-flight and point spread function techniques. In terms of applied activity, ¹⁸F-FDG had a median of 314.55 ± 62.69 MBq.

2.3 Image Segmentation

The CT and PET image sequences were imported into the IntelliSpace Discovery platform version 3.0 software developed by Philips Healthcare in Best. This software has the capability to simultaneously present the corresponding layers of PET and CT images through the utilization of the “Research Oncology Suite” plug-in unit, as seen in Appendix S1.1. Utilizing the metabolic information provided by PET images and the information provided by CT images, researchers defined the high FDG uptake region of the lesion as the volume of interest (VOI). This approach is predicated on the observation that the majority of HGSOC lesions exhibit a cystic-solid composition. Typically, the cystic component contains a minimal number of tumor cells, whereas the solid component harbors a substantial number of tumor cells, which demonstrate elevated metabolic activity on PET images. Using the solid component of the lesion to generate indicators of spatial heterogeneity may more accurately capture the discrepancies among tumor cells.

The identification of peritoneal metastatic lesions in HGSOC was based on the anatomical abdominopelvic 9-region method [30, 31]. By intersecting two sagittal planes and two transverse planes, the abdomen is anatomically divided into nine regions. The two sagittal planes split the abdominal cavity into three equal segments. The two transverse planes are the lowest margin of costal arch and the plane of anterior superior iliac spine. The umbilical region is designated area 0 (AR0). Starting from the right upper part to the right middle part, the regions are numbered AR1-AR8 clockwise [32], as seen in Appendix S1.2. The nine regions were sequentially assessed for each patient. If a lesion was identified in a particular region, the VOI was delineated accordingly. In other words, each patient had 2–10 VOIs (including primary lesion). Each VOI was required to be greater than 5 × 5 × 5 mm. We involved two professional researchers (SBH and ZDG, who have interpreted PET/CT images for 12 and 15 years, respectively) to evaluate all the regions, and disagreements were resolved by a senior researcher (GZY, with 23 years of experience). SBH outlined the VOIs of all lesions. To assess the intra- and inter-group consistency, SBH and ZDG repeated the segmentation process independently on 30 randomly chosen cases after a lapse of 1 month.

2.4 Using Conventional Features to Generate Spatial Heterogeneity Indicators

Eight conventional features were recorded for each VOI in the pathological area. Four CT features included CT value, minor axis length, major axis length, and pixel number. Among the PET features were total lesion glycolysis and three standard uptake values (SUV): mean SUV, peak SUV, and maximum SUV. The following 13 statistics were calculated for each conventional measurement at the patient level: mean, standard deviation, median, quartile deviation, mode, variance, range, standard error of the sample mean, uncorrected sum of squares, corrected sum of squares, coefficient of variation, kurtosis, and skewness. Accordingly, for each patient, 13 × 8 = 104 spatial heterogeneity indicators were generated. The detailed method can be found in Appendix S1.3.

2.5 Using Texture Features to Generate Spatial Heterogeneity Indicators

The relevant functions in the Computational Environment for Radiological Research package were used to calculate CT texture feature parameters [33]. The radiomic texture features were defined in accordance with the Image Biomarker Standardization Initiative guidelines [34]. Similar voxels were subsequently classified into clusters based on the similarity of their textural features. Initially, the CT images underwent rescaling to utilize 256 gray levels, followed by discretization using a bin width of 32. Subsequently, the gray-level co-occurrence matrix was computed for each voxel within the VOI [35]. Four unrelated features, namely Energy, Entropy, Contrast, and Homogeneity, also known as Haralick texture features, were extracted at the voxel level. The Gaussian Mixture model clustering algorithm [36] was used to convert voxels into subregions with homogeneous textures. The textural dissimilarities between subregion pairs were calculated as the Euclidean distance [19] and displayed as a dissimilarity matrix. A spatial heterogeneity indicator called cluster site entropy (cSE) was calculated using the dissimilarity matrix [37]. The two-dimensional histogram was designated the group diversity matrix (GDM), from which two spatial heterogeneity indicators, cluster diversity (cluDev) and cluster standard deviation (cludiss), were calculated [32].

2.6 Immunohistochemical Staining and Scoring

Following surgery, all patients underwent pathological section preparation, a process carried out by the pathology department at our hospital. The pathological sections were preserved in 10% formaldehyde and encased in paraffin. In order to ensure precise quantification of the degree of immunohistochemical staining, we scanned the immunohistochemical sections to generate digital whole slide images using a Pannoramic MIDI scanner (3DHISTECH, Hungary). Furthermore, we calculated the immunohistochemical scores using Aipathwell (Servicebio, China). This software utilizes artificial intelligence technology, specifically deep learning principles, to train algorithms using extensive data sets and incorporate them into automated image analysis processes. The outcomes were determined through various stages, including tracking, color selection, operation, and analysis [38]. The software initially identified and outlined the region of interest on the tissue automatically. Subsequently, utilizing whole slide images, it autonomously conducted positive assessments, identified the nucleus, and extended the cytoplasmic boundaries. It then computed the quantity and size of weak, moderate, and strong positive cells, in addition to various parameters including integrated optical density and tissue area. Ultimately, it systematically calculated the test region at a high level of magnification using the initial data and algorithmic formulas, culminating in the generation of an analytical report. To assess the link between immunohistochemical scores and spatial heterogeneity, Spearman's rank correlation coefficient was used.

2.7 Platinum Resistance Feature Selection and Modeling

Three groups of prediction models were established: conventional, heterogeneity, and integrated models. The collected data used to establish platinum resistance models is seen in Appendix S1.4. For each group of models, H₂O Automatic Machine Learning (H2O AutoML) was used for feature selection and modeling (https://docs.h2o.ai/h2o/latest-stable/h2o-docs/automl.html). This method can perform multiple algorithms simultaneously, like GBM, discriminative random field (DRF), extremely randomized trees (XRT), generalize linear model (GLM), Stacked Ensemble and DeepLearning. Following the completion of automated learning, the H2O system will generate the top 10 models exhibiting the highest performance according to their respective parameters. This research sought to identify the optimal model by evaluating the area under the curve (AUC) metric. The training process adopts 5-fold cross validation. The training set and validation set were from the Nanhu Campus (NH) and Huaxiang Campus (HX) of our hospital, respectively.

2.8 Statistical Analysis

The Kolmogorov–Smirnov test was utilized to assess the normality of the data. Normalized descriptive data were shown as means ± standard deviations, while non-normalized data was shown as medians, upper quartile (1st Qu.), and lower quartile quartiles (3rd Qu). t-tests were used for comparisons between groups composed of normally distributed data, Mann–Whitney U tests for non-normally distributed data, and chi-square tests for enumeration data. The intraclass correlation coefficient (ICC) [39] was used to evaluate intra- and inter-group consistency.

Spatial heterogeneity indicator extraction was carried out utilizing the software MATLAB R2022a (MathWorks, USA), while platinum resistance feature selection and modeling were performed using R software (version 4.1.0). Key packages utilized in the analysis included “corrplot,” “lattice,” “rms,” “ggplot2,” and “h₂o.” Statistical significance was determined at a p < 0.05 (two-tailed). The data and analysis scripts for this study can be accessed on GitHub at https://github.com/beibei22008/platinum-resistance_HGSOC.

3.1 Patient Characteristics

Initially, 1697 patients performed PET/CT scans and were diagnosed with suspected malignant ovarian masses. After screening, 286 patients were ultimately included in this study, as shown in Figure 1. There were no significant differences (p > 0.05) in the clinical information and lesion characteristics between participants in the training set and those in the validation set. Tables 1 and 2 provided summaries of the clinical information and lesions' characteristics of the participants.

3.2 Prediction Models for Platinum Resistance

The heterogeneity indicators showed good consistency both in intra- and inter-group, with ICC > 0.75. The collected data were used to establish three groups of platinum resistance models. After H₂O AutoML, the GBM model proved to be the optimal model for conventional, heterogeneity, and integrated models. A comparison of the AUC and area under curve of precision recall (AUCPR) for the different learning algorithms in each group models is shown in Figure 2. The AUC and AUCPR metrics of the GBM model surpassed those of the other algorithmic models. Consequently, the GBM model was selected as the optimal model within each group for subsequent comparisons.

We then compared the chosen optimal models in each group (conventional, heterogeneity, and integrated model). The order of the AUC in the validation set was as follows: integrated (0.928, 95% CI: 0.872, 0.984) > heterogeneity (0.904, 95% CI: 0.842, 0.966) > conventional (0.790, 95% CI: 0.696, 0.885) model. The integrated model had the best prediction performance Figure 3. showed AUC, AUCPR, and calibration curves for the selected optimal models within each group.

Focusing on the integrated model, the optimal model was GBM_grid_1_AutoML_2_20230923_170056_ Model _4. The detailed evaluation indicators of the models were summarized in Table 3. Cross-Validation showed the mean accuracy was 0.831 ± 0.045, and the mean specificity was 0.820 ± 0.110.

To conduct a more detailed examination of the internal feature composition and the relative significance of each feature within the models, we employed Variable Importance Heatmaps and Variable Importance Plots. The Variable Importance Heatmap (Figure 4A–C) displayed the percentage values of the top ten most significant features across multiple models, facilitating the selection of optimal models. And the Variable Importance Plot (Figure 4D–F) illustrated the relative importance of the most critical features within each model, with each feature's importance normalized on a scale from 0 to 1.

The learning curve plot (Figure 5A–C) illustrated the error metric's dependence on the learning progress of each model. Gains/Lift graphs (Figure 5D–F) assessed the improvement and overall accuracy of the model in predicting platinum resistance.

3.3 Correlation Analysis

Based on the Variable Importance Plots of the integrated model, we selected the ten features with the highest scores for the next correlation analysis. Eighty-six patients have performed the p53 immunohistochemical staining. Spearman's correlation analysis revealed a significant positive association between HU_Kurtosis and the immunohistochemical score for p53 (ρ = 0.718, p < 0.01), indicating a strong correlation. Among the 84 patients who underwent Ki-67 immunohistochemical staining, cSE exhibited the strongest positive association between the immunohistochemical level for Ki-67 with correlation coefficient ρ = 0.759 (p < 0.01), as illustrated in Figure 6.