Deep learning of pretreatment ascites cytopathology for platinum-resistance risk stratification in advanced epithelial ovarian cancer

Abstract

Background

Platinum resistance is a major determinant of poor outcome in advanced epithelial ovarian cancer, yet reliable predictors available before treatment initiation remain scarce. Ascitic fluid is commonly obtained during diagnostic work-up and directly reflects the peritoneal tumour microenvironment, but its cytomorphological information has not been systematically exploited for treatment-response prediction.

Methods

We present OVCAP, a multi-scale deep-learning framework that analyses pretreatment ascites cytology whole-slide images to estimate platinum-resistance risk. The study included 438 patients with FIGO stage IIIB–IV epithelial ovarian cancer. Model performance was evaluated in one internal and two independent external validation cohorts. Attention-guided cytopathology review was performed to identify high-risk morphologic patterns, and integrated single-cell RNA sequencing analyses were used to characterise the underlying biological features.

Results

OVCAP achieved area under the receiver operating characteristic curve (ROC-AUC) values of 0.894, 0.863, and 0.828 in the internal and two independent external validation cohorts, respectively, and outperformed the KELIM score (AUC 0.619). Attention-guided cytopathology review identified recurrent high-risk morphologic patterns in resistant disease: epithelial cytoplasmic vacuolization and interaction-rich malignant aggregates accompanied by immune and mesothelial cells. Integrated single-cell analyses linked these phenotypes to membrane remodelling, lipid reprogramming, hypoxia-associated stress signalling, and reinforced adhesion and immunoregulatory networks.

Conclusion

These findings support pretreatment ascites cytology as a clinically accessible substrate for early risk stratification before first-line platinum-based therapy.

Graphical abstract

Introduction

Ovarian cancer remains one of the most lethal gynaecologic malignancies, with most patients presenting at advanced stages due to ineffective screening and non-specific early symptoms [1,2]. In this setting, first-line management typically involves either primary debulking surgery (PDS) followed by platinum-based chemotherapy or neoadjuvant chemotherapy (NACT) followed by interval debulking surgery (IDS) [3]. While randomized trials have reported similar overall survival between PDS and NACT-IDS, prognosis in practice depends on both achieving complete cytoreduction and responding to platinum-based chemotherapy [4,5]. Accordingly, the choice of initial treatment must be individualized rather than applied uniformly.

A major source of uncertainty is inter-patient heterogeneity in response to NACT and platinum-based chemotherapy [6]. Clinically, platinum chemosensitivity is commonly defined by the platinum-free interval (PFI): recurrence ≥6 months after completion of treatment is considered platinum-sensitive, whereas earlier recurrence or progression during therapy is considered platinum-resistant or refractory. This information is inherently retrospective [2,7]. Existing prognostic tools, including radiologic assessment, CA-125 kinetics, chemotherapy response score (CRS) on IDS specimens, and BRCA/HRD-related assays, provide useful information but are typically available only after several treatment cycles or require surgical tissue, limiting their utility for guiding decisions at NACT initiation [[8], [9], [10], [11], [12]]. There is therefore a clear need for early, minimally invasive predictors of platinum response in advanced epithelial ovarian cancer.

Malignant ascites is a hallmark of advanced and recurrent ovarian cancer, reflecting extensive peritoneal dissemination, aggressive tumour biology, and poor prognosis [13]. Ascites forms a tumour-promoting microenvironment containing malignant cells, stromal and immune populations, soluble mediators, and extracellular vesicles that contribute to metastasis and chemoresistance [[14], [15], [16]]. Therapeutic paracentesis and cytopathologic examination of ascitic fluid are routine and can be serially sampled with minimal invasiveness, making ascites an attractive biospecimen for longitudinal, minimally invasive assessment [17,18]. However, conventional ascitic cytopathology is largely qualitative and diagnostic, focusing on the presence of malignant cells rather than capturing and quantifying morphologic features that might encode chemosensitivity.

Advances in artificial intelligence (AI)-based computational pathology have demonstrated that morphologic images encode rich information about tumour biology, molecular status, prognosis, and treatment response [[19], [20], [21]]. Deep-learning models applied to whole-slide images can detect malignancy, infer tumour origin across multiple cancer types [[22], [23], [24]]. In ovarian cancer, however, AI studies on ascites whole-slide cytopathology have remained largely centred on identifying malignant cells, with little attention to predicting pretreatment platinum response or platinum-resistance risk [23]. Because ascitic cytopathology retains multicellular tumour clustering without the full architecture of tissue, it may encode treatment-relevant biology beyond simple malignant-cell detection [25]. Thus, the potential of ascitic cytopathology as a quantitative, AI-enabled biomarker of platinum response remains largely unexplored.

Here, we developed OVCAP (OVarian Cytopathology model for Assessing Platinum response), a deep-learning model that analyses pretreatment ascites cytology whole-slide images to predict platinum-resistance risk in advanced epithelial ovarian cancer patients receiving first-line platinum-based chemotherapy. Across multiple cohorts, OVCAP enabled pretreatment risk stratification before systemic therapy. We further linked high-risk model predictions to distinct cytomorphological and microenvironmental features in ascites, providing biological context for the image-derived signal and suggesting mechanisms that may underlie platinum failure.

Methods

Study design and patient cohorts

This retrospective multi-institutional study included 438 adult women (≥18 years) with histologically confirmed FIGO stage IIIB-IV epithelial ovarian cancer who underwent pretreatment ascitic fluid collection and cytopathological evaluation at two institutions across three clinical sites: the Xuhui campus of Fudan University Shanghai Cancer Center (FUSCC), the Pudong campus of FUSCC, and Northern Jiangsu People’s Hospital. The study was conducted in accordance with the Declaration of Helsinki and approved by the Medical Ethics Committee of Fudan University Shanghai Cancer Center (approval no. 2108241-25), with corresponding approval obtained at participating centres as required. The requirement for informed consent was waived because of the retrospective nature of the study and the use of deidentified data. Patients were enrolled between 2019 and 2023.

Eligible patients received first-line platinum-based chemotherapy, either in the neoadjuvant setting followed by interval debulking surgery or in the adjuvant setting after primary debulking surgery. Patients were excluded if (1) cytology slides were unavailable or inadequate for digital analysis, (2) tumour histology was non-epithelial ovarian cancer, or (3) key clinical variables were missing (Fig. 1). All included patients had complete baseline clinicopathological information and evaluable platinum-response outcomes.

Fig. 1.

Study design and cohort allocation.

Flowchart of patient selection from the Fudan University Shanghai Cancer Center (FUSCC) Xuhui and Pudong cohorts and the Northern Jiangsu People’s Hospital (NJPH) cohort, and assignment of 438 patients with FIGO stage IIIB–IV disease to the training, internal validation, and two external validation cohorts.

Dataset allocation

Among the 438 eligible patients, cases from the Xuhui campus of FUSCC (n=367) were randomly divided into a training set (n=254, 69.2%) and an internal validation set (n=113, 30.8%) using stratified random sampling according to platinum-response status to preserve class balance. Two independent external validation cohorts were obtained from the Pudong campus of FUSCC (external validation cohort 1, n=51) and Northern Jiangsu People’s Hospital (external validation cohort 2, n=20). Internal and external validation cohorts were held out throughout model development and were not used for hyperparameter tuning.

Clinical characteristics and outcome definition

Clinical variables included age at diagnosis, BMI, FIGO stage, histological subtype, treatment information, and platinum-response status. Full baseline characteristics are summarized in Supplementary Table S1. Platinum resistance was defined as radiologic recurrence or progression within 6 months after completion of platinum-based chemotherapy, or as progressive disease during platinum-based treatment. Platinum sensitivity was defined as radiologic recurrence occurring at least 6 months after completion of platinum-based chemotherapy. Platinum-free interval (PFI) was calculated from the date of the last platinum dose to the date of radiologically documented recurrence or progression. Radiologic progression or recurrence was assessed according to RECIST v1.1. GCIG CA-125 criteria were used as supportive evidence where applicable.

Whole slide images preprocessing and patch extraction

Ascitic cytology processing and whole-slide images (WSIs) acquisition procedures are detailed in the Supplementary Methods. WSIs were preprocessed to identify foreground regions and generate patch coordinates for downstream feature extraction. Foreground regions were segmented using the HEST tissue-versus-background segmentation model with a confidence threshold of 0.4 [26]. Patch coordinates were then generated at the magnification and patch size required by each foundation model (Supplementary Table S2). Patches containing more than 50% background were excluded from subsequent analysis.

Feature extraction with foundation models

Patch-level feature extraction was performed using the unified encoder interface provided by TRIDENT [27]. To identify the optimal feature extractor for ascites cytology, we benchmarked 15 encoders, including 14 pathology foundation models and an ImageNet-pretrained ResNet50 baseline (Supplementary Table S2). For each encoder, patch embeddings were extracted from WSIs using the encoder-specific magnification and patch size recommended by TRIDENT. Each patch was represented by a fixed-dimensional feature vector defined by the corresponding pretrained model, and all features were L2-normalized prior to aggregation. All encoder-specific preprocessing settings were applied according to the TRIDENT implementation to ensure fair comparison across encoders.

Multi-scale attention-based multiple instance learning

We employed an attention-based multiple instance learning (MIL) framework in which each WSI was treated as a bag of patch embeddings. Encoder-specific patch embeddings were first projected into a shared latent space through a two-layer feed-forward network before being input into the TransMIL backbone [28]. In the implementation used here, the input embeddings were mapped from the encoder-specific dimension to 512 dimensions and then to a 256-dimensional latent representation. The TransMIL backbone used Nyström attention to approximate self-attention and incorporated a positional encoding generator to capture spatial relationships among patches. The latent representations were processed using two transformer layers with a hidden dimension of 256, followed by WSI-level classification using the class token.

To capture image information at multiple spatial scales, we constructed three parallel streams operating on patch-, cell-, and thumbnail-level representations. Each stream independently produced a probability score for platinum resistance. The outputs from the three streams were concatenated and combined through a logistic regression layer for late fusion. The final prediction was computed as

$$P=\sigma \left({\beta }_0+{\beta }_{1}logit\left(p_{patches}\right)+{\beta }_{2}logit\left(p_{cell}\right)+{\beta }_{3}logit\left(p_{thumbnail}\right)\right)$$

where $\sigma$ denotes the sigmoid function, $p_{patches}$, $p_{cell}$, $p_{thumbnail}$ are the stream-level predicted probabilities, and ${\beta }_0$, ${\beta }_1$, ${\beta }_2$, ${\beta }_3$ are learned fusion coefficients. The fitted fusion model was subsequently applied unchanged to the internal and external validation cohorts.

Model training and hyperparameter selection

Model development was conducted exclusively within the training cohort. To identify the optimal encoder and training configuration, hyperparameter optimization was performed using grid search combined with stratified three-fold cross-validation at the patient level to avoid data leakage. Candidate configurations were generated over multiple hyperparameters including encoder type, number of sampled patches, learning rate, and batch size. For each configuration, models were trained across the three cross-validation folds and repeated using 10 random seeds to account for stochastic variation during optimization. Validation performance was evaluated using the area under the receiver operating characteristic curve (ROC-AUC). The encoder-hyperparameter configuration achieving the highest mean validation ROC-AUC across folds and seeds was selected as the final configuration (Supplementary Table S3). The final model was then retrained on the full training cohort using the selected configuration. All model training and inference were performed using PyTorch (v2.3.1) on NVIDIA A100 80GB GPUs.

Model interpretation and attention visualization

Attention scores were extracted from the trained TransMIL model for each WSI. For a slide containing n patches, the model produced attention weights representing the relative importance of each patch for the final prediction. High-attention patches, defined as the top 1% ranked by attention score, were extracted from platinum-resistant cases for cytopathologist review (Supplementary Methods) to highlight the most influential regions contributing to model predictions.

Single-cell RNA sequencing data acquisition and processing

A publicly available single-cell RNA sequencing (scRNA-seq) dataset of ovarian cancer ascites was obtained from GSA-Human under accession code PRJCA005422 [14]. The deposited AnnData object was imported into Scanpy (v1.4.6), and only high-grade serous ovarian cancer (HGSOC) ascites samples were retained for downstream analysis (Supplementary Table S4).

Quality-control metrics included the number of detected genes, total UMI counts, and mitochondrial transcript percentage. Cells with fewer than 200 detected genes, more than 6000 detected genes, or greater than 20% mitochondrial content were excluded. Putative doublets were identified and removed prior to downstream analysis using Scrublet (v0.2.3) with default parameters. Raw counts were then normalized to 10,000 counts per cell and log-transformed.

Highly variable genes were identified using the Seurat v3 method with batch_key = "Samples", and the top 5000 highly variable genes were retained. Batch correction was performed using scvi-tools (v1.2.2.post2) with Samples specified as the batch variable (n_layers=2, n_latent =30). UMAP visualization and Leiden clustering were then performed on the scVI latent space, and cell types were annotated using CellTypist (v1.6.3) followed by manual curation based on canonical marker genes. Marker genes used for major cell-type annotation and differentially expressed genes for major cell clusters are provided in Supplementary Tables S5 and S6, respectively.

Gene ontology and pathway activity analyses

Functional program activity in ascites-derived malignant cells was evaluated using complementary gene set-based approaches. For Gene Ontology (GO) analysis, GO Biological Process gene sets were obtained. Per-cell gene set activity scores were then calculated using AUCell-based enrichment implemented in omicverse (v1.7.7). To compare functional programs associated with platinum response, per-cell AUCell score matrices were compared between platinum-resistant and platinum-sensitive epithelial cells using group-wise differential analysis implemented in Scanpy (scanpy.tl.rank_genes_groups).

In parallel, signalling pathway activity was inferred using the PROGENy[29] model implemented in decoupler (v2.2.1). PROGENy weights were used to estimate the activity of 14 canonical signalling pathways, including TNFα, Hypoxia, Androgen, TGF-β, EGFR, VEGF, JAK-STAT, PI3K, MAPK, NFκB, p53, TRAIL, WNT, and Estrogen signalling, at single-cell resolution. Pathway activity scores were computed for each cell and subsequently compared between platinum-resistant and platinum-sensitive groups.

Metabolite-mediated communication analysis

Metabolite-mediated cell-cell communication was inferred using MEBOCOST (v1.2.2)[30] based on cell type-resolved single-cell transcriptomic profiles. Significant metabolite-sensor interactions between sender and receiver cell populations were identified using permutation-based testing with multiple-testing correction. Communication burden and interaction structure were summarized across annotated cell types and visualized using event count plots and network diagrams.

Ligand-receptor interaction analysis

Ligand-receptor-mediated cell-cell communication was inferred using CellChat (v1.6.1) [31]. Overexpressed ligands and receptors were identified within annotated cell populations, and communication probabilities were estimated based on the law of mass action with permutation-based significance testing. Inferred interactions were aggregated at the signalling pathway level and organized into major functional categories, including adhesion, extracellular matrix, immunoregulatory signalling, and growth factor pathways. Communication networks were visualized as cell-cell interaction graphs.

Statistical analysis

All statistical analyses were performed using Python (v3.10.0) and R (v4.2.2). Classification performance was evaluated using ROC-AUC, PR-AUC, accuracy, recall, specificity, and F1-score. For model selection experiments involving cross-validation, performance metrics were summarized by averaging across folds and random seeds. For held-out internal and external validation cohorts, 95% confidence intervals were estimated by bootstrap resampling of case-level predictions with 10,000 iterations. Concordance among single-scale and fusion predictors was assessed using Kendall’s τ. Other statistical analyses, including group comparisons and regression analyses, are described in the corresponding figure legends where applicable. All statistical tests were two-sided.

Results

Study cohort and dataset allocation

Our study included 438 patients in the final analysis (Fig. 1). All patients had FIGO stage IIIB-IV epithelial ovarian cancer, underwent pretreatment ascitic cytological evaluation, and had complete clinical information available from two institutions across three clinical sites. Exclusion criteria included unavailable cytology slides, non-epithelial histology, or missing key clinical variables. Among the eligible patients, the primary dataset was randomly split into training and internal validation cohorts at a 7:3 ratio, comprising 254 and 113 patients, respectively. Two independent external cohorts (n=51 and n=20) served as additional validation sets.

Baseline characteristics were generally balanced across cohorts. The median age was comparable across cohorts (approximately 56-57 years; p=0.899), and BMI showed no significant difference among groups (p=0.499). KELIM scores demonstrated a statistically significant difference between cohorts (p=0.002), whereas the distribution of FIGO stage III versus IV disease remained balanced (p=0.376). Most patients presented with serous histology (>85% across all cohorts). Importantly, the proportion of platinum-resistant disease remained consistent across groups (approximately 20%; p=0.916), supporting comparability of outcome distribution across cohorts. Baseline characteristics are summarized in Supplementary Table S1.

Ascites cytology-based deep learning framework for platinum response prediction

We developed a deep-learning framework leveraging pretreatment ascites cytology whole-slide images (WSIs) to predict platinum-resistance risk in patients with ovarian cancer (Fig. 2). Ascitic samples were collected before treatment, processed into H&E-stained cytology slides, and digitized for computational analysis (Fig. 2A). Each WSI was tiled into non-overlapping patches, and patch-level representations were extracted using pretrained foundation model encoders to ensure robust feature learning (Supplementary Table S2) [27]. These features were subsequently aggregated within an attention-based multiple instance learning (MIL) architecture.

Fig. 2.

Overview of ascites cytology acquisition and the OVCAP framework.

(A) Pretreatment ascites collection, cytology slide preparation, and generation of H&E-stained whole-slide images (WSIs). (B) Schematic of the OVCAP pipeline. Patch-, cell-, and thumbnail-level features were extracted from cytology WSIs to generate branch-specific risk scores, which were integrated by logistic regression–based late fusion to obtain the final OVCAP score. Model development and validation were performed across the Fudan University Shanghai Cancer Center (FUSCC) Xuhui and Pudong cohorts and the Northern Jiangsu People’s Hospital (NJPH) cohort.

We employed a TransMIL backbone augmented with Nyström attention to efficiently capture global contextual dependencies and a position-perception-enhanced guidance module to improve long-range spatial awareness and discriminative capacity [28]. To further exploit complementary cytological cues at different image resolutions, we constructed a multi-scale architecture incorporating patch-level, cell-level, and thumbnail-level predictors. Each stream generated an independent prediction probability, and these outputs were then integrated through logistic regression–based late fusion to derive the final WSI-level prediction (Fig. 2B).

Benchmarking pathology foundation models for cytology prediction

Given the limited maturity and external benchmarking of cytology-specific foundation models, we systematically evaluated whether leading histopathology foundation models could be adapted for ascites cytology. We benchmarked 15 pretrained encoders, including 14 representative pathology foundation models and an ImageNet-pretrained ResNet50 baseline (Supplementary Table S2) [27,[32], [33], [34], [35]]. Benchmarking was performed within the training dataset using repeated stratified cross-validation under multiple random seeds, with the best-performing configuration retained for each encoder. Performance was assessed using area under the receiver operating characteristic curve (ROC-AUC), area under the precision–recall curve (PR-AUC), accuracy, recall, specificity, and F1-score. As summarized in Supplementary Table S3, the framework demonstrated stable performance across encoders, confirming robustness. Notably, Prov-GigaPath[32] achieved the highest mean AUC (0.868), followed by H-Optimus and UNI-based models, whereas conventional CNN baselines showed inferior performance, highlighting the superiority of large-scale transformer-based and self-supervised pathology models. Detailed benchmarking distributions are provided in Fig. 3.

Fig. 3.

Benchmarking of pathology foundation models for ascites cytology prediction.

Performance comparison of 15 pretrained encoders, including 14 pathology foundation models and an ImageNet-pretrained ResNet50 baseline, for pretreatment ascites cytology prediction. Bar plots show area under the receiver operating characteristic curve (AUC), area under the precision–recall curve (AUPRC), F1 score, accuracy (ACC), recall, and specificity for each encoder. Bars indicate mean performance across repeated stratified cross-validation runs, and error bars denote the standard deviation across runs.

Multi-scale fusion improves prediction performance and generalizability across cohorts

During model development, the proposed multi-scale framework demonstrated reliable predictive capability, with clear class separation and limited misclassification compared with the single-scale models (Supplementary Fig. S1). We next assessed its generalizability across an internal validation cohort and two independent external cohorts (Fig. 4). The three single-scale predictors exhibited only weak pairwise concordance (Kendall’s τ=0.09-0.17), while each showed only moderate correlation with the fusion model, indicating that they captured complementary cytological signals (Fig. 4A). Consistently, standardized logistic regression analyses confirmed that patch-, cell-, and thumbnail-level features each contributed independently to response prediction, with the fusion model demonstrating the strongest effect size (Fig. 4B).

Fig. 4.

Performance comparison of single-scale and fusion models across validation cohorts.

(A) Kendall’s τ correlation matrix showing concordance among patch-, thumbnail-, cell-, and fusion-level predictors. (B) Standardized odds ratios from logistic regression showing the independent contribution of each modality to the fusion model. (C) Receiver operating characteristic (ROC) curve of the KELIM score for predicting platinum resistance. (D) Confusion matrices of the fusion model in the internal validation cohort, external validation cohort 1, and external validation cohort 2. (E) Comparison of predictive performance between the fusion and single-scale models across the three validation cohorts, including area under the ROC curve (ROC-AUC), accuracy, area under the precision–recall curve (PR-AUC), specificity, and F1-score. Error bars indicate 95% confidence intervals estimated by bootstrap resampling.

To contextualize the image-derived signature relative to clinical determinants, univariable and multivariable logistic regression analyses were performed in the training cohort, internal validation cohort, and external validation cohort 1. Across these datasets, OVCAP score remained the only consistently significant clinical predictor, whereas age, FIGO stage, neoadjuvant chemotherapy, and maintenance therapy contributed minimally (Supplementary Fig. S2A-C), suggesting that the cytology-based computational biomarker provides prognostic value largely independent of conventional clinical variables. Logistic regression was not performed in external validation cohort 2 due to limited sample size (n=20), which would yield unstable effect estimation. Furthermore, comparison with KELIM demonstrated only modest discrimination for this serum kinetic biomarker (AUC=0.619), markedly inferior to the fusion model (Fig. 4C), underscoring the additional value of cytological computational phenotyping.

Across the internal and external cohorts, the fusion framework maintained favourable classification behaviour with fewer misclassifications relative to the single-scale models (Fig. 4D). Bootstrap resampling (10,000 iterations) further confirmed statistical robustness, with consistently superior ROC-AUC, PR-AUC, accuracy, specificity, and F1-score compared with patch-, cell-, and thumbnail-level predictors. In the internal validation cohort, the fusion model achieved an ROC-AUC of 0.894, PR-AUC of 0.711, accuracy of 0.858, specificity of 0.956, and F1-score of 0.547 (Fig. 4E). Superior performance was consistently preserved in both external cohorts, with ROC-AUCs of 0.863 and 0.828, accompanied by improved PR-AUC, accuracy, and specificity compared with the single-scale models (Fig. 4E). Together, these findings demonstrate that the multi-scale fusion framework is robust, reproducible, and shows consistent performance across internal and external validation cohorts.

Phenotype 1

Epithelial vacuolization and metabolic reprogramming

To biologically contextualize model-derived attention signals, we extracted the top 1% highest-attention cytology patches from each platinum-resistant case and subjected them to independent review and consensus adjudication by two senior cytopathologists. Notably, we identified two recurrent high-risk cytomorphological phenotypes.

The first phenotype (Fig. 5A), characterized by epithelial cells with clear, translucent cytoplasm and prominent cytoplasmic vacuolization, often arranged as cohesive multicellular clusters with indistinct cell borders, prompted us to interrogate epithelial cell states at single-cell resolution (Supplementary Fig. S3). Gene Ontology-based scoring revealed coordinated enrichment of pathways involved in tight junction organisation, glycolipid and cholesterol transport, carbohydrate derivative trafficking, and organic hydroxy compound transport. These programs were enriched in resistant epithelial cells and were consistent with altered membrane organisation, vesicular trafficking, and lipid transport (Fig. 5B).

Fig. 5.

Biologically interpretable high-risk image phenotypes associated with platinum resistance.

(A) Representative high-attention cytology patches from platinum-resistant cases with high OVCAP scores, highlighting the epithelial vacuolization phenotype. (B) Dot plot of enriched Gene Ontology (GO) biological processes in resistant versus sensitive epithelial cells. (C) Differential PROGENy pathway activity in resistant and sensitive epithelial cells, highlighting TNFα, hypoxia, and androgen signalling. (D) Differential metabolite-centred activities inferred by MEBOCOST in resistant and sensitive epithelial cells. (E) Representative high-attention cytology patches from platinum-resistant cases showing interaction-rich malignant aggregates. (F) Schematic of the resistant ascites tumour microenvironment, including epithelial tumour cells, cancer-associated mesothelial cells (MCs), cancer-associated fibroblasts (CAFs), tumour-associated macrophages (TAMs), dendritic cells (DCs), natural killer (NK) cells, T and B lymphocytes, cancer-associated adipocytes (CAAs), and secreted factors, extracellular vesicles, and metabolites. (G) Ro/e heatmap showing differential enrichment of annotated cell states in resistant and sensitive ascites. (H) Heatmap of differential ligand–receptor signalling pathways between resistant and sensitive ascites. (I) Bar plot summarizing metabolite-mediated cell–cell communication (mCCC) events across sender and receiver cell populations in resistant ascites.

Consistent with these epithelial programs, PROGENy-based pathway inference revealed a distinct stress-adaptive signalling landscape in resistant epithelial cells, characterized by significantly higher TNFα-associated inflammatory signalling, a pronounced hypoxic response, and elevated androgen signalling compared with sensitive samples (Fig. 5C).

Further extending these observations to metabolic states, we applied MEBOCOST to infer metabolite-centred pathway activities at the single-cell level. Resistant samples exhibited preferential enrichment of lipid- and eicosanoid-associated metabolites, including 5-HETE, prostaglandin derivatives, and acyl-CoA-linked intermediates, together with stress-associated nitrogen and redox metabolic signatures. In contrast, sensitive samples displayed higher engagement of mitochondrial cofactor–related metabolites such as lipoamide, along with glucuronic acid and carbohydrate-associated intermediates, indicating a more physiological metabolic routing. These findings support a model in which resistant epithelial cells adopt a stress-adaptive, inflammatory lipid-dominated metabolic configuration within cohesive multicellular aggregates, whereas sensitive cells retain relatively preserved mitochondrial and carbohydrate metabolism (Fig. 5D).

Phenotype 2

Interaction-rich tumour microenvironment

The second phenotype (Fig. 5E) was characterized by highly malignant epithelial cell aggregates forming compact, spheroid-like clusters closely accompanied by immune and mesothelial cells. We next investigated whether this interaction-rich cytological phenotype was associated with distinct cellular composition and intercellular communication programs in ascites. Cytology and schematic representation revealed a structurally cohesive and interaction-rich ecosystem within resistant ascites, consisting of epithelial tumour cells closely integrated with mesothelial-derived elements, cancer-associated fibroblasts, macrophages, dendritic cells, NK cells, T and B lymphocytes, adipocyte-like components, and abundant secreted factors and extracellular vesicles (Fig. 5F).

To quantitatively evaluate how these cellular populations were differentially represented across clinical groups, we next calculated Ro/e-based enrichment scores for each annotated cell state in resistant versus sensitive ascites (Fig. 5G). A spectrum of immune populations showed preferential over-representation in resistant samples, including multiple CD8⁺ T-cell states (ANXA2⁺, CX3CR1⁺, GZMK⁺, CCR7⁺), CD4⁺ T-cell subsets (ANXA1⁺, PDCD1⁺, CX3CR1⁺, CCR7⁺, FOXP3⁺), monocyte/macrophage populations (FCGR3A⁺ and CD14⁺ subsets, as well as proliferative macrophage states), hematopoietic stem-like cells, follicular B cells, and proliferative cell compartments. In contrast, these populations generally exhibited lower enrichment in sensitive samples. These data indicate that resistant ascites harbour a more prominently expanded immune cellular landscape at the compositional level.

To determine whether such compositional expansion was accompanied by rewiring of intercellular communication, we next compared ligand-receptor signalling programs between clinical phenotypes (Fig. 5H). Resistant ascites displayed coordinated reinforcement of adhesion-, extracellular matrix-, and immunoregulatory signalling axes, highlighted by selective upregulation of ICAM, ITGB2, CCL, GALECTIN, MHC-II, ANNEXIN, ADGRE5, MK, FN1, LAMININ, COLLAGEN, and APP pathways, predominantly engaging macrophage, dendritic cell, T-cell, NK-cell, mesothelial, and fibroblast compartments. In contrast, MHC-I, MIF, CD99, CD45, RESISTIN, SPP1, CLEC, and SELPLG signalling modules were attenuated in resistant samples and relatively preserved in sensitive ascites.

Consistently, MEBOCOST-based inference of metabolite-mediated multicellular communication circuits demonstrated a substantially increased communication burden and dominant macrophage-dendritic-mesothelial-epithelial hubs in resistant ascites (Fig. 5I). Together, these observations indicate that platinum resistance is associated with epithelial metabolic-transport reprogramming and a structurally and immunologically organized, interaction-rich ascites ecosystem, providing biologically coherent interpretability for the model’s attention-derived image signatures.

Illustrative examples of potential clinical integration

To conceptually demonstrate how OVCAP might be incorporated into future clinical workflows, we selected representative cases from the 438 patients already included in our training and validation cohorts to simulate its application in pretreatment decision contexts. No additional samples were collected for this illustration. In a case assigned a low predicted risk of platinum resistance, the model highlighted sparse tumour burden with relatively preserved cytological organisation (Fig. 6A). Conversely, in a case assigned a high predicted risk, dense malignant clusters and interaction-rich cytological features were emphasized (Fig. 6B). These simulated examples are intended only to illustrate potential decision-support integration, rather than representing prospective deployment or real-world treatment guidance, and warrant validation in future prospective settings.

Fig. 6.

Illustrative examples of OVCAP-based pretreatment risk assessment.

(A) Representative case with a low OVCAP score and lower predicted risk of platinum resistance. (B) Representative case with a high OVCAP score and higher predicted risk of platinum resistance.

Discussion

In this study, we developed OVCAP, a multi-scale deep learning framework that uses pretreatment ascites cytology whole-slide images to estimate platinum-resistance risk in advanced epithelial ovarian cancer. Across internal and independent external validation cohorts, OVCAP showed consistent discriminatory performance before treatment initiation. Beyond predictive performance, attention-guided pathology review and integrative single-cell analyses linked high-risk model outputs to recurrent cytomorphological and microenvironmental features, providing biological context for the image-derived predictions. Although platinum resistance is strongly associated with poor outcomes, reliable biomarkers available before first-line treatment remain limited. Existing indicators, including radiologic response assessment, CA-125 kinetics, and chemotherapy response scoring, are often obtained only after treatment has started or require surgical specimens [[8], [9], [10], [11], [12]]. In contrast, malignant ascites is readily accessible in advanced disease and directly reflects the peritoneal tumour niche [13]. Recent AI studies in ascites cytology have focused primarily on malignancy detection, showing that ovarian cancer can be identified from ascitic whole-slide images, but they did not address pretreatment platinum resistance, treatment stratification, or the biological interpretation of resistance-associated image phenotypes [23]. By extending computational pathology to pretreatment ascites cytology in advanced epithelial ovarian cancer, our study identifies a clinically accessible and minimally invasive substrate for early response stratification before systemic therapy.

Two principal findings emerge from this work. The first is that pretreatment ascites cytology contains reproducible morphological information associated with platinum resistance, and that this information is distributed across multiple spatial levels [36]. Ascitic cytology represents a fluid and heterogeneous disease compartment in which tumour cells appear as isolated cells, multicellular aggregates, and mixed populations with immune and mesothelial cells [37]. This organisation is unlikely to be captured adequately at a single scale. Consistent with this biology, the multi-scale fusion model outperformed the individual patch-, cell-, and thumbnail-level models, suggesting that platinum-resistance–associated signals in ascites are encoded across complementary levels of cytological structure.

The second principal finding is that the model-derived image signatures were biologically coherent. We identified two recurrent high-risk phenotypes in platinum-resistant ascites: epithelial cytoplasmic vacuolization and interaction-rich malignant aggregates accompanied by immune and mesothelial cells. In the accompanying single-cell analyses, the vacuolization phenotype was associated with pathways related to membrane organisation, vesicular trafficking, glycolipid and cholesterol transport, and stress-adaptive signalling, together with lipid- and eicosanoid-associated metabolic programs [38,39]. By contrast, the aggregate-rich phenotype was associated with broader enrichment of immune cell states and reinforced ligand-receptor and metabolite-mediated communication networks centred on macrophage, dendritic, mesothelial, and epithelial compartments [40,41]. Together, these observations suggest that platinum resistance in ascites is associated with both tumour-cell-intrinsic stress adaptation and microenvironmental rewiring.

These findings also support the potential translational value of ascites cytopathology as a pretreatment risk-stratification biomarker. Because ascitic samples are routinely obtained in advanced ovarian cancer and can be assessed before systemic therapy, OVCAP may offer a practical means of early risk stratification without requiring additional invasive tissue acquisition. In this setting, computational analysis of ascitic cytology could complement existing clinical and molecular variables by providing morphology-based information that is available at the start of treatment. More broadly, our results support the concept that ascitic cytology is not only diagnostically informative, but also predictive of treatment-relevant tumour biology.

Several limitations warrant consideration. First, the retrospective design may introduce selection bias, highlighting the need for prospective, multi-centre validation of OVCAP. Relatedly, because the cohort was predominantly composed of high-grade serous ovarian carcinoma, the applicability of OVCAP to low-grade serous ovarian carcinoma and other less common histologic subtypes requires further validation in larger histology-diverse cohorts. Second, variability in ascites cytology preparation and digitization workflows across institutions may affect model generalizability, emphasizing the importance of standardized protocols and large-scale validation. Third, although the recurrent high-risk cytomorphological phenotypes were biologically contextualized by single-cell transcriptomic analyses, their mechanistic basis requires further investigation. Future studies integrating functional experiments, ascites biochemical profiling, and experimental models based on ascites-derived tumour cells are needed to clarify how epithelial vacuolization and interaction-rich malignant aggregates relate to platinum resistance. Finally, integration with complementary biomarkers such as circulating tumour DNA or proteomics may further enhance predictive accuracy and clinical utility.

Conclusion

In summary, OVCAP demonstrates that pretreatment ascites cytology contains clinically meaningful and biologically interpretable signals associated with platinum resistance in advanced epithelial ovarian cancer. Rather than relying on post-treatment indicators, this work highlights a minimally invasive, treatment-start substrate for early risk stratification. With further prospective validation and standardisation, computational analysis of ascites cytology may provide a useful addition to pretreatment decision-support in ovarian cancer.

Data and code availability

The public single-cell RNA sequencing dataset analyzed in this study is available from GSA-Human under accession code PRJCA005422. Raw clinical data are not publicly available due to patient privacy concerns and ethical restrictions. Deidentified ascites cytology whole-slide images used in this study are available from the corresponding author upon reasonable request and subject to institutional approval. The code for the OVCAP framework and analysis pipelines is publicly available at https://github.com/pigudog/OVCAP. Additional information required to reanalyze the data reported in this study is available from the corresponding author upon reasonable request.

Funding

This work was supported by the National Natural Science Foundation of China (grant Nos. 82272898, 82203723, and 82471932), the Three-Year Action Plan to Promote Clinical Skills and Clinical Innovation Capacity of Municipal Hospitals by the Shanghai Shenkang Hospital Development Center (grant No. SHDC2020CR5003-001), and the CSCO-CYH Oncology Research Fund (Y-Young2024-0297).

CRediT authorship contribution statement

Yangyang Zhang: Writing – review & editing, Writing – original draft, Methodology, Investigation, Formal analysis, Conceptualization. Xiaochun Wan: Writing – review & editing, Writing – original draft, Methodology, Investigation, Conceptualization. Yongqi Chen: Writing – review & editing, Writing – original draft, Methodology, Formal analysis. Jianbo Xu: Writing – review & editing, Writing – original draft, Methodology. Weijie Wang: Writing – review & editing, Investigation. Haiming Li: Writing – review & editing, Investigation. Zhihao Zhang: Writing – review & editing, Investigation. Yi-Hua Luo: Writing – review & editing, Investigation. Liu Wang: Writing – review & editing, Supervision, Conceptualization. Xingzhu Ju: Writing – review & editing, Supervision, Conceptualization. Xiaohua Wu: Writing – review & editing, Supervision, Conceptualization. Zilong Wang: Writing – review & editing, Supervision, Methodology, Conceptualization. Bo Ping: Writing – review & editing, Supervision, Conceptualization. Qinhao Guo: Writing – review & editing, Supervision, Conceptualization.

Declaration of competing interest

The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.