Skip to main content
Download PDF

Development of model for identifying homologous recombination deficiency (HRD) status of ovarian cancer with deep learning on whole slide images

Journal of Translational Medicine volume 23, Article number: 267 (2025) Cite this article

Abstract

Background

Homologous recombination deficiency (HRD) refers to the dysfunction of homologous recombination repair (HRR) at the cellular level. The assessment of HRD status has the important significance for the formulation of treatment plans, efficacy evaluation, and prognosis prediction of patients with ovarian cancer.

Objectives

This study aimed to construct a deep learning-based classifier for identifying tumor regions from whole slide images (WSIs) and stratify the HRD status of patients with ovarian cancer (OC).

Methods

The deep learning models were trained on 205 H&E-stained sections which contained 205 ovarian cancer patients, 64 were found to have HRD status while 141 had homologous recombination proficiency (HRP) status from two institutions Memorial Sloan Kettering Cancer Center (MSKCC) and Zhongda Hospital, Southeast University. The framework includes tumor regions identification by UNet + + and subtypes of ovarian cancer classifier construction. Referring to the EasyEnsemble, we classified the HRP patients into three distributed subsets. These three subsets of HRP patients were combined with the HRD patients to establish three new training groups for subsequent model construction. The three models were integrated into a single model named Ensemble Model.

Results

The UNet + + algorithm segmented tumor regions with 81.8% accuracy, 85.9% recall, 83.8% dice score and 68.3% IoU. The AUC of the Ensemble Model was 0.769 (Precision = 0.800, Recall = 0.727, F1-score = 0.762) in the study. The most discriminative features between HRD and HRP comprised S_mean_dln_obtuse_ratio, S_mean_dln_acute_ratio and mean_Graph_T-S_Betweenness_normed.

Conclusions

The models we constructed enables accurate discrimination between tumor and non-tumor tissues in ovarian cancer as well as the prediction of HRD status for patients with ovarian cancer.

Introduction

Ovarian cancer is one of the three major reproductive tumors that seriously threaten women’s health around the world [1]. Meanwhile, although the mortality rate of ovarian cancer has decreased in recent years, its mortality rate ranks second among gynecological malignant tumors, following malignant tumor of uterine [2]. Early symptoms of ovarian cancer are often atypical and there is a lack of effective screening methods [3]. Most ovarian cancer patients are diagnosed at an advanced stage of the disease, with distant spread or metastasis of the tumor, and the prognosis is poor usually [4]. Hence, the early detection of ovarian cancer and enhancing survival rates for patients have emerged as a pivotal research focus worldwide. For patients with ovarian cancer, the predominant treatment option consists of a combination of surgery and chemotherapy primarily [5]. Regrettably, the majority of ovarian cancer patients may experience disease recurrence and progress to a platinum-resistant state, resulting in a bleak prognosis [6].

Homologous recombination deficiency (HRD) refers to the dysfunction of homologous recombination repair (HRR) at the cellular level, which can be caused by many factors such as germline or somatic mutations of HRR related genes and epigenetic inactivation [7]. Under normal circumstances, in the face of DNA damage in cells, a variety of repair methods including HRR in the human body will make up for DNA damage. However, DNA damage cannot be properly repaired by HRR if HRD is present. The accumulation of DNA double-strand breaks leads to genomic and chromosomal instability. Poly ADP-ribose polymerase (PARP) inhibitors are a class of anti-tumor drugs. PARP inhibitors can inhibit the function of PARP enzyme to block the DNA damage repair pathway, thereby promoting the apoptosis of tumor cells and playing a stronger anti-tumor effect for patients with HRD, especially for those who suffer from ovarian cancer [8]. Therefore, the assessment of HRD status has the important significance for the formulation of treatment plans, efficacy evaluation, and prognosis prediction of patients with ovarian cancer. The clinical detection methods of HRD can be classified into three categories: HRR-related gene mutation detection, genomic scar and mutation pedigree analysis, and functional detection of HRD [9]. Although these methods can determine the HRD status of patients, they are costly and time-consuming, making them less widespread, especially in underdeveloped regions. Hence, there is an imperative requirement for a more effective and expedient HRD detection system that can offer a quick and accurate determination of the HRD status for patients with ovarian cancer.

Pathology represents a crucial aspect of cancer patient information and are instrumental in providing clinical practitioners with necessary support for patient management, outcome assessment, and treatment decision-making during the clinical practice [10, 11]. In recent times, the field of digital pathology has advanced significantly due to progress in science and technology. The foundation of digital pathology involves the use of a pathology scanner to scan solid pathological section images. The digitally captured images are then stored and interpreted in a digital file format to produce digital pathological whole slide images (WSIs). This approach provides clinical researchers with the advantage of more precise and convenient diagnostic tools, which can be used for diagnosis and analysis.

The continuous advancement of deep learning technology has introduced novel prospects for pathological diagnosis and treatment and has significantly contributed to the expansion of digital pathology. Currently, deep learning technology is primarily employed in three distinct areas of digital pathology: (1) screening and diagnosis, (2) prediction of treatment response, and (3) prognosis prediction [12,13,14,15]. In previous investigations, we have effectively devised a system that utilizes deep learning algorithms to extract pathological features and predict the prognosis of patients afflicted with breast cancer [16]. Furthermore, we have also conducted investigations involving digital pathology techniques, which have enabled us to differentiate between and accurately stratify the risk associated with basal cell carcinoma [17]. These deep learning applications have demonstrated remarkable efficacy towards achieving more precise and efficient diagnosis and developing personalized treatment strategies based on a patient’s specific condition.

Similarly, the field of ovarian cancer research has also witnessed significant progress because of the ingenious integration of deep learning and digital pathology techniques. In the study conducted by Nero et al., a publicly available weakly supervised deep learning-based method was employed, which succeeded in accurately predicting the BRCA gene mutation status of ovarian cancer patients using digital pathology [18]. Our team is inspired by the achievements and envisages developing a novel model aimed at identifying the HRD status of ovarian cancer patients through the application of deep learning technologies that leverage informative pathology data.

Hence, the objective of our research is to develop two advanced models that employ a synergistic combination of UNet++ [19]generated tumor images and nuclear features based on the histopathologic images, which contain morphology, texture, spatial arrangement of tumor cell nucleus, interrelationships between tumor cells and other cells to differentiate ovarian cancer and stratify the HRD status of ovarian cancer.

Materials and methods

General dataset introduction

We aimed to investigate a cohort of 205 patients who had been definitively diagnosed with ovarian cancer based on postoperative pathology and HRD status, including dataset 1 (D1): 183 patients from Memorial Sloan Kettering Cancer Center (MSKCC) [20] and dataset 2 (D2): 22 patients from Zhongda Hospital, Southeast University. To effectively examine and analyze the data, all patients in the cohort had provided digital scan pathological slides alongside comprehensive clinical data. In this study, patients with no clear evidence to confirm HRD status were recognized as homologous recombination proficiency (HRP) status.

In this study, WSIs were obtained from the routine hematoxylin and eosin (H&E) diagnostic tissue slides and the quality of each slide was evaluated by two professional pathologists to ensure its reliability. Moreover, one slice of the highest quality was meticulously chosen for each patient, which was subsequently included in this study.

HRD detection

The following are the HRD detection methods used in the 2 cohorts included in this study. In D1, the research team from MSKCC employed clinical sequencing to infer HRD status, focusing specifically on variants in genes associated with the HRD DNA damage response (DDR), such as BRCA1 and BRCA2, as well as those characteristic of distinct mutational subtypes enriched for tandem duplications and foldback inversions [20, 21]. In D2, DNA extraction and HRD detection were conducted in collaboration with the laboratory personnel from Nanjing Simcere Company. The HRD assessments in this research employed next-generation sequencing (NGS) technology to identify mutations in BRCA1/2 and other homologous recombination repair (HRR)-related genes. Additionally, the HRD score was computed by evaluating loss of heterozygosity (LOH), large-scale state transitions (LST), and telomeric allelic imbalance (TAI), thereby providing a comprehensive determination of the tumor’s HRD status. Tumors were classified as exhibiting HRD status when either of the following criteria was met: (1) an HRD score of ≥ 40, or (2) the presence of pathogenic variants in BRCA1/2 genes. Conversely, tumors not meeting these criteria were designated as HRP. The HRD Score model based on the Chinese population data can reflect the HRD status of the Chinese population compared with the Myriad myChoice® CDx product approved by Food and Drug Administration (FDA) [22].

Tumor region identification and nuclear segmentation

The pathological sections were meticulously annotated using QuPath software by two experienced pathologists who categorized the following areas: tumor and other tissue. In cases where inconsistencies arose between the two sets of annotations, a senior pathologist reviewed the designated areas and made the final determination. The segmentation process utilized a sliding window operation of 512 pixels × 512 pixels to extract image blocks from full-field digital slices captured at 20x. These image blocks were input to the UNet + + model for training, and the patches in D1 data were randomly divided into 80% for training and 20% for internal testing to evaluate the performance of the model. Ultimately, the predicted image blocks were combined to obtain the complete segmented image of the ovarian cancer tumor area.

The segmentation model developed in this research was trained on D1 and excels at distinguishing the tumor region and other tissue by making accurate pixel-level predictions in the test images. To ascertain the model’s efficacy, we evaluated the performance of the model in terms of pixel precision, recall, intersection over union (IoU), and dice score. Moreover, the model we have built was applied to the pathological sections in D2 for automatic identification and extraction of images containing the tumor area.

The deep convolutional neural network method called Hover-Net is a powerful tool utilized in this study for both instance segmentation and classification of nuclei. The model was originally trained on a publicly available dataset known as CoNSeP [23]. In 2024, Pan et al. using the Hover-Net model to automatically segment different cell types in a tumor region, successfully predicted the survival of patients with lung adenocarcinoma based on pathological features [24]. In addition, in a previous study by our team, it was also by segmentation of different types of cells in the tumor region by the Hover-Net model that successfully classified the risk of basal cell carcinoma [17]. Therefore, in this research, we employed the Hover-Net model to segment cells present within the tumor area, classify each nucleus by type, and calculate pertinent morphology, texture, and spatial features.

Feature extraction

In this study, the histopathological features within the tumor region were carefully extracted and analyzed. The aim of the present study was to closely examine histopathological features in the tumor region by extracting nuclear features and tumor microenvironment (TME) profiles. A total of 240 dimensions were obtained from our feature extraction process, encompassing texture, shape, spatial arrangement of cell nucleus characteristics and TME features.

Nuclear morphological/texture features (46 descriptors): These features focused on quantifying nuclear texture and shape using different measurements.

Nuclear spatial arrangement and TME features (194 descriptors): This set of characteristics are intended to capture the difference of nuclear topology and spatial structural relationship between tumor nuclei in corresponding conditions.

Feature selection

This study employed three different feature selection methods, namely, minimum redundancy maximum relevance (mRMR), Least Absolute Shrinkage and Selection Operator (LASSO), and Random Forest (RF), to identify the optimal features that most effectively distinguish between the two classes of ovarian cancer (HRD and HRP) in the modeling set. To avoid overfitting and the dimensionality curse, we established a maximum feature limit of 10 for the machine learning classifier. This limitation was implemented to address issues related to the generalization performance of the classifier, which can be compromised when the number of features exceeds a threshold relative to the number of training samples.

Classifier construction and evaluation

In the research, a significant disparity was observed in the number of HRD and HRP patients, indicating an imbalance in the data. In response to the observed data imbalance, referring to the EasyEnsemble [25], we classified the HRP patients into three distributed subsets. Correspondingly, these three subsets of HRP patients were combined with the HRD patients to establish three new training groups (ND1, ND2, ND3) for subsequent model construction.

To accomplish the aims of this study, we utilized four distinct machine learning classifiers: Logistic Regression (LR), k-Nearest Neighbors (KNN), Support Vector Machine (SVM), and Random Forest (RF). These classifiers were combined with three different feature selection methods, yielding a total of 12 possible combinations. For each of the ND1-3 datasets, the data were randomly partitioned into an 80% training group and a 20% internal validation group to assess model performance. The most effective feature selection method and machine learning classifier were identified for each of the ND1-3 datasets based on the area under the receiver operating characteristic curve (AUC) derived from all images in the training group, using five-fold cross-validation, and these optimal combinations were designated as Model 1, Model 2, and Model 3, respectively. Subsequently, the three models (Model 1–3) generated from the previously established ND1-3 were integrated into a single model named Ensemble Model. This Ensemble Model was then applied to the external testing group (D2) to evaluate its performance, assessed through measures such as AUC, F1-score, recall, and specificity. The entire workflow of our research is illustrated in Fig. 1.

Fig. 1
figure 1

Flow chart

Full size image

Results

Patients characteristics of two datasets

The study cohort consisted of 205 ovarian cancer patients, 64 were found to have HRD status while 141 had HRP status, with a mean age of 63.81 years (SD: 10.79; range 37–89 years). Meanwhile, 97.6% patients were diagnosed in the advanced stages of the disease. All patients received surgical treatment, with 142 (69.3%) patients undergoing neadjuvant chemotherapy-interal debulking surgery (NACT-IDS) and 63 (30.7%) patients received the primary debulking surgery (PDS). Only 47 patients (22.9%) were administered PARP inhibitor treatment. The clinical and pathological features of the patients are presented in Table 1.

Table 1 Summary of clinical and pathological features of the patients
Full size table

Tumor regions identification and the performance of subtype classifier

The UNet + + algorithm segmented tumor regions with 81.8% accuracy, 85.9% recall, 83.8% dice score and 68.3% IoU (Table 2). Regarding to the differentiation classifier, the performance of the 12 combination of feature selection and classifier schemes in terms of classification performance for differentiation subtypes HRD and HRP of ovarian cancer are summarized in Table 3. The combination of RF and RF analysis achieved the best performance on WSIs in distinguishing HRD and HRP of ovarian cancer in ND1 (AUC = 0.817 ± 0.052, precision = 0.751 ± 0.117, recall = 0.736 ± 0.070, F1-score = 0.734 ± 0.050). The combination with the best performance was RF and RF in ND2 (AUC = 0.815 ± 0.043, precision = 0.814 ± 0.167, recall = 0.736 ± 0.026, F1-score = 0.763 ± 0.069). Meanwhile, the combination with the best performance was RF and RF in ND3 (AUC = 0.838 ± 0.051, precision = 0.780 ± 0.160, recall = 0.764 ± 0.065, F1-score = 0.757 ± 0.066). The results are presented in Fig. 2A-C. Moreover, our model demonstrated excellent performance in both the internal testing groups (Fig. 2D-E, Table 1S) and external testing group (Fig. 3A-C, Table 2S), indicating its reliability.

Table 2 The performance of tumor region segmentation model
Full size table
Table 3 The performance of the 12 combination of feature selection and classifier schemes for discriminating subtypes from ovarian cancer in training groups
Full size table
Fig. 2
figure 2

Performance of the subtype classifier on training groups. Training group in ND1-3 (A-C). Internal testing group in ND1-3 (D-F)

Full size image
Fig. 3
figure 3

The Receiver Operating Characteristic Curve (ROC) of the subtype classifiers on external testing groups. Model 1 (A), Model 2 (B), Model 3 (C), Ensemble Model (D)

Full size image

Subsequently, we integrated the three models constructed (Models 1–3) into the single model called the Ensemble Model and evaluated its effect using the external testing group. The Ensemble Model has the great performance for differentiation subtypes HRD and HRP of ovarian cancer. The AUC of the Ensemble Model was 0.769 (Precision = 0.800, Recall = 0.727, F1-score = 0.762) in the external testing group. The results are shown in Table 4; Fig. 3D.

Table 4 The performance of Ensemble Model
Full size table

The discriminative features of differentiation classifier

In the training groups, there were 10-dimension discriminative features used to construct the differentiation classifier (Fig. 4). We ranked the features by the contribution of each feature to the classification target, which is obtained by the feature importances function of the RF (Table 3S).

Fig. 4
figure 4

Discriminative top 10 features of the subtype classifier of ovarian cancer. Model 1 (A), Model 2 (B), Model 3 (C)

Full size image

A total of three features were present in all three models including S_mean_dln_obtuse_ratio, S_mean_dln_acute_ratio and mean_Graph_T-S_Betweenness_normed. In this study, we discovered that HRD patients exhibited a distinctive spatial arrangement of tumor-infiltrating lymphocyte nuclei, demonstrating greater dispersion in comparison to non-HRD patients. Furthermore, the distance between tumor cells and necrotic cells was found to be significantly closed in the HRD patient group. These findings suggest that the HRD status of ovarian cancer may have a significant impact on the tumor microenvironment and cell-to-cell interactions.

Discussion

For patients diagnosed with epithelial ovarian cancer, the standard approach for initial treatment involves a comprehensive evaluation of the patient’s condition, followed by surgery such as NACT-IDS and PDS. Additionally, adjuvant chemotherapy with platinum-based drugs is administered after surgery [26]. Recently, PARP inhibitors have emerged as a novel class of targeted therapeutics with significant therapeutic achievements in the maintenance treatment of ovarian cancer, particularly in patients with HRD status [27]. The detection of HRD may enable the use of PARP inhibitors in approximately 50% of ovarian cancer patients who present with this molecular subtype, offering a promising avenue for targeted therapy [28].

One of the current challenges faced by clinicians is the rapid identification of ovarian cancer patients with HRD status, enabling them to derive maximum benefit from therapeutic application of PARP inhibitors. It is noteworthy that there is currently no globally standardized protocol for the detection of HRD status. Presently available technologies face challenges in directly determining HRD status of ovarian cancer patients through clinical information and histopathological features. Gene testing is commonly employed to identify HRD status, with products such as Myriad myChoice® CDx and FoundationFocusTM CDx BRCA LOH currently FDA-approved for clinical use. Nonetheless, these tests are time-consuming and expensive, presenting major drawbacks.

In recent years, the increasing development of deep learning technology has highlighted its potential in the field of medical image processing [29]. There is growing enthusiasm for the use of deep learning to address clinical challenges and improve patient outcomes. Histopathologic scanning of cancer tissue typically involves capturing images at extremely high magnifications, resulting in file sizes that can reach into the billions of pixels. As a result, these pathological images contain a vast amount of tumor microenvironment data, which is crucial in studying the progression of cancer. Nero et al. utilized an openly available deep learning-based weakly supervised method called clustering-constrained-attention multiple-instance learning (CLAM) to construct a diagnostic model capable of identifying the BRCA1/2 gene status in ovarian cancer patients from WSIs that had undergone H&E staining in 2022 [18]. However, the model developed by the aforementioned team demonstrated suboptimal performance, with a validation ROC AUC of 0.59, indicating the necessity for further optimization. In contrast, our team has developed a more excellent model based on WSIs through deep learning approaches. By incorporating nuclear morphological/texture features along with nuclear spatial arrangement and TME features, our model exhibits superior performance in discriminating HRD status in ovarian cancer patients. The determination of HRD status holds significant clinical implications for therapeutic outcome prediction in ovarian cancer and compared to single-gene mutation status assessment, HRD status evaluation offers enhanced potential for patient stratification and optimization of therapeutic decision-making [27, 30, 31].

Afterwards, in 2023, Bourgade’s research team proposed a novel deep learning classifier with the ability to predict the BRCA mutation status from Whole Slide Images (WSIs) of ovarian cancer patients, following H&E staining [32]. This approach exhibited significantly improved performance relative to previous methods, highlighting the potential of deep learning in digital pathology to achieve high levels of precision and efficiency in cancer diagnosis. In addition, Wang’s team developed two deep learning methods capable of predicting the therapeutic response to Bevacizumab in ovarian cancer and the microsatellite instability status of ovarian cancer patients on WSIs, without the need for pathologists to conduct detailed image annotation [33]. This innovative approach offers significant benefits, including the potential to enhance the efficiency of cancer diagnosis and treatment, while also reducing the workload of pathologists in analyzing WSIs.

In this investigation, we created two models based on digitized H&E-stained histopathological images of ovarian cancer patients. The first model was developed to identify the tumor region in ovarian cancer patients, while the second model was designed to discriminate whether the patient of ovarian cancer with HRD status or not. The features obtained from the WSIs are closely related to various aspects of the image, such as nuclear orientation, shape, and texture. Our study reveals for the first time the most representative pathological features in HRD and HRP ovarian cancer, which may help distinguish the clinicopathological diagnosis in the future.

Notwithstanding the significant findings of this study, some limitations are worth noting. Firstly, the utilization of patient image data obtained from various datasets without controlling for typical confounding variables such as pathological section quality, diverse patient demographics, and distinct treatment measures. These variables can significantly impact both the quality of images and the subsequent feature analysis procedures, thus affecting the overall accuracy and reliability of the results. Consequently, future studies should aim to address these confounding variables to improve the quality of image data and accurately reflect the patient population under investigation. Furthermore, it is imperative to recognize that the current study is constrained by a relatively limited sample size. To augment the robustness and generalizability of the developed model, it is crucial to expand the patient cohort in subsequent investigations. As our research advances, we are committed to the ongoing refinement and optimization of the constructed models. We have commenced the integration of additional external validation cohorts from diverse geographical regions and healthcare institutions. This expansion will encompass a more comprehensive array of patients and clinical data, including overall survival, status of recurrence, tumor pathological types and so on. Such an approach will facilitate a more thorough evaluation of the model’s performance across a wider range of patient demographics and clinical scenarios, thereby enhancing its applicability and reliability in real-world settings.

Conclusion

In summary, our study demonstrates that employing deep learning-based pathological image analysis can effectively perform routine pathological tasks. Our models we constructed enables accurate discrimination between tumor and non-tumor tissues in ovarian cancer as well as the prediction of HRD status for patients with ovarian cancer. In future research, the adoption of deep learning can potentially identify further pathological relevant features and facilitate the construction of more precise models, thus contributing to advancing healthcare and promoting medical knowledge more rapidly.

Data availability

The following information was supplied regarding data availability: Data in the Memorial Sloan Kettering Cancer Center (MSKCC) is available from the Synapse (Sage Bionetworks) (Accession Code: syn25946117). The data that support the findings of this study are available from the corresponding author upon reasonable request.

References

  1. Peres LC, Sinha S, Townsend MK, Fridley BL, Karlan BY, Lutgendorf SK, Shinn E, Sood AK, Tworoger SS. Predictors of survival trajectories among women with epithelial ovarian cancer. Gynecol Oncol. 2020;156:459–66.

    Article  PubMed  Google Scholar 

  2. Siegel RL, Giaquinto AN, Jemal A. Cancer statistics, 2024. CA Cancer J Clin. 2024;74:12–49.

    Article  PubMed  Google Scholar 

  3. Lalwani N, Prasad SR, Vikram R, Shanbhogue AK, Huettner PC, Fasih N. Histologic, molecular, and cytogenetic features of ovarian cancers: implications for diagnosis and treatment. Radiographics. 2011;31:625–46.

    Article  PubMed  Google Scholar 

  4. Reid BM, Permuth JB, Sellers TA. Epidemiology of ovarian cancer: a review. Cancer Biol Med. 2017;14:9–32.

    Article  CAS  PubMed  PubMed Central  Google Scholar 

  5. Stewart C, Ralyea C, Lockwood S. Ovarian Cancer: an Integrated Review. Semin Oncol Nurs. 2019;35:151–6.

    Article  PubMed  Google Scholar 

  6. Giornelli GH. Management of relapsed ovarian cancer: a review. Springerplus. 2016;5:1197.

    Article  PubMed  PubMed Central  Google Scholar 

  7. Lau CH, Seow KM, Chen KH. The Molecular mechanisms of actions, effects, and clinical implications of PARP inhibitors in epithelial ovarian cancers: a systematic review. Int J Mol Sci 2022, 23.

  8. Li J, Li Q, Zhang L, Zhang S, Dai Y. Poly-ADP-ribose polymerase (PARP) inhibitors and ovarian function. Biomed Pharmacother. 2023;157:114028.

    Article  CAS  PubMed  Google Scholar 

  9. Ledermann JA, Drew Y, Kristeleit RS. Homologous recombination deficiency and ovarian cancer. Eur J Cancer. 2016;60:49–58.

    Article  CAS  PubMed  Google Scholar 

  10. Flanigan JS, Silkensen SL, Wolf NG. Global Health Pathology Research: purpose and funding. Clin Lab Med. 2018;38:21–35.

    Article  PubMed  Google Scholar 

  11. Moch H, Blank PR, Dietel M, Elmberger G, Kerr KM, Palacios J, Penault-Llorca F, Rossi G, Szucs TD. Personalized cancer medicine and the future of pathology. Virchows Arch. 2012;460:3–8.

    Article  CAS  PubMed  Google Scholar 

  12. Wei JW, Suriawinata AA, Vaickus LJ, Ren B, Liu X, Lisovsky M, Tomita N, Abdollahi B, Kim AS, Snover DC, et al. Evaluation of a Deep Neural Network for Automated Classification of Colorectal Polyps on histopathologic slides. JAMA Netw Open. 2020;3:e203398.

    Article  PubMed  PubMed Central  Google Scholar 

  13. Zhang X, Gleber-Netto FO, Wang S, Martins-Chaves RR, Gomez RS, Vigneswaran N, Sarkar A, William WN Jr., Papadimitrakopoulou V, Williams M, et al. Deep learning-based pathology image analysis predicts cancer progression risk in patients with oral leukoplakia. Cancer Med. 2023;12:7508–18.

    Article  PubMed  PubMed Central  Google Scholar 

  14. Wang K, Ren Y, Ma L, Fan Y, Yang Z, Yang Q, Shi J, Sun Y. Deep learning-based prediction of treatment prognosis from nasal polyp histology slides. Int Forum Allergy Rhinol. 2023;13:886–98.

    Article  PubMed  Google Scholar 

  15. Zhou G, Zheng J, Chen Z, Hu D, Li S, Zhuang W, He Z, Lin G, Wu B, Zhang W, et al. Clinical significance of tumor-infiltrating lymphocytes investigated using routine H&E slides in small cell lung cancer. Radiat Oncol. 2022;17:127.

    Article  CAS  PubMed  PubMed Central  Google Scholar 

  16. Lu C, Romo-Bucheli D, Wang X, Janowczyk A, Ganesan S, Gilmore H, Rimm D, Madabhushi A. Nuclear shape and orientation features from H&E images predict survival in early-stage estrogen receptor-positive breast cancers. Lab Invest. 2018;98:1438–48.

    Article  CAS  PubMed  PubMed Central  Google Scholar 

  17. Lan X, Guo G, Wang X, Yan Q, Xue R, Li Y, Zhu J, Dong Z, Wang F, Li G, et al. Differentiation and risk stratification of basal cell carcinoma with deep learning on histopathologic images and measuring nuclei and tumor microenvironment features. Skin Res Technol. 2024;30:e13571.

    Article  PubMed  PubMed Central  Google Scholar 

  18. Nero C, Boldrini L, Lenkowicz J, Giudice MT, Piermattei A, Inzani F, Pasciuto T, Minucci A, Fagotti A, Zannoni G et al. Deep-learning to predict BRCA mutation and survival from Digital H&E slides of epithelial ovarian Cancer. Int J Mol Sci 2022, 23.

  19. Zhou Z, Siddiquee MMR, Tajbakhsh N, Liang J. UNet++: redesigning skip connections to exploit Multiscale features in image segmentation. IEEE Trans Med Imaging. 2020;39:1856–67.

    Article  PubMed  Google Scholar 

  20. Boehm KM, Aherne EA, Ellenson L, Nikolovski I, Alghamdi M, Vázquez-García I, Zamarin D, Long Roche K, Liu Y, Patel D, et al. Multimodal data integration using machine learning improves risk stratification of high-grade serous ovarian cancer. Nat Cancer. 2022;3:723–33.

    Article  CAS  PubMed  PubMed Central  Google Scholar 

  21. Zehir A, Benayed R, Shah RH, Syed A, Middha S, Kim HR, Srinivasan P, Gao J, Chakravarty D, Devlin SM, et al. Mutational landscape of metastatic cancer revealed from prospective clinical sequencing of 10,000 patients. Nat Med. 2017;23:703–13.

    Article  CAS  PubMed  PubMed Central  Google Scholar 

  22. Scaglione GL, Pignata S, Pettinato A, Paolillo C, Califano D, Scandurra G, Lombardo V, Di Gaudio F, Pecorino B, Mereu L et al. Homologous recombination Deficiency (HRD) Scoring, by means of two different shallow whole-genome sequencing pipelines (sWGS), in Ovarian Cancer patients: a comparison with myriad MyChoice assay. Int J Mol Sci 2023, 24.

  23. Graham S, Vu QD, Raza SEA, Azam A, Tsang YW, Kwak JT, Rajpoot N. Hover-Net: simultaneous segmentation and classification of nuclei in multi-tissue histology images. Med Image Anal. 2019;58:101563.

    Article  PubMed  Google Scholar 

  24. Pan X, Feng S, Wang Y, Chen J, Lin H, Wang Z, Hou F, Lu C, Chen X, Liu Z, et al. Spatial distance between tumor and lymphocyte can predict the survival of patients with resectable lung adenocarcinoma. Heliyon. 2024;10:e30779.

    Article  PubMed  PubMed Central  Google Scholar 

  25. Liu XY, Wu J, Zhou ZH. Exploratory undersampling for class-imbalance learning. IEEE Trans Syst Man Cybern B Cybern. 2009;39:539–50.

    Article  PubMed  Google Scholar 

  26. Kuroki L, Guntupalli SR. Treatment of epithelial ovarian cancer. BMJ. 2020;371:m3773.

    Article  PubMed  Google Scholar 

  27. Ray-Coquard I, Pautier P, Pignata S, Pérol D, González-Martín A, Berger R, Fujiwara K, Vergote I, Colombo N, Mäenpää J, et al. Olaparib plus Bevacizumab as First-Line maintenance in Ovarian Cancer. N Engl J Med. 2019;381:2416–28.

    Article  CAS  PubMed  Google Scholar 

  28. Konstantinopoulos PA, Ceccaldi R, Shapiro GI, D’Andrea AD. Homologous recombination Deficiency: exploiting the fundamental vulnerability of Ovarian Cancer. Cancer Discov. 2015;5:1137–54.

    Article  CAS  PubMed  PubMed Central  Google Scholar 

  29. Barragán-Montero A, Javaid U, Valdés G, Nguyen D, Desbordes P, Macq B, Willems S, Vandewinckele L, Holmström M, Löfman F, et al. Artificial intelligence and machine learning for medical imaging: a technology review. Phys Med. 2021;83:242–56.

    Article  PubMed  PubMed Central  Google Scholar 

  30. Ray-Coquard I, Leary A, Pignata S, Cropet C, González-Martín A, Marth C, Nagao S, Vergote I, Colombo N, Mäenpää J, et al. Olaparib plus Bevacizumab first-line maintenance in ovarian cancer: final overall survival results from the PAOLA-1/ENGOT-ov25 trial. Ann Oncol. 2023;34:681–92.

    Article  CAS  PubMed  Google Scholar 

  31. DiSilvestro P, Banerjee S, Colombo N, Scambia G, Kim BG, Oaknin A, Friedlander M, Lisyanskaya A, Floquet A, Leary A, et al. Overall survival with maintenance olaparib at a 7-Year Follow-Up in patients with newly diagnosed Advanced Ovarian Cancer and a BRCA mutation: the SOLO1/GOG 3004 Trial. J Clin Oncol. 2023;41:609–17.

    Article  CAS  PubMed  Google Scholar 

  32. Bourgade R, Rabilloud N, Perennec T, Pécot T, Garrec C, Guédon AF, Delnatte C, Bézieau S, Lespagnol A, de Tayrac M, et al. Deep learning for detecting BRCA mutations in high-Grade Ovarian Cancer based on an innovative tumor segmentation method from whole slide images. Mod Pathol. 2023;36:100304.

    Article  PubMed  Google Scholar 

  33. Wang CW, Lee YC, Lin YJ, Firdi NP, Muzakky H, Liu TC, Lai PJ, Wang CH, Wang YC, Yu MH, et al. Deep learning can predict Bevacizumab therapeutic effect and microsatellite instability directly from Histology in Epithelial Ovarian Cancer. Lab Invest. 2023;103:100247.

    Article  CAS  PubMed  Google Scholar 

Download references

Acknowledgements

We are grateful to the staff in Biobank of Zhongda Hospital Affiliated to Southeast University for technical assistance.

Funding

This study was supported by Jiangsu Provincial Commission of Health (M2022016), Postgraduate Research & Practice Innovation Program of Jiangsu Province (SJCX23_0090), Zhongda Hospital Affiliated to Southeast University, Jiangsu Province High-Level Hospital Pairing Assistance Construction Funds (zdyyxy07), National clinical key discipline construction funds(czxm-zk-40).

Author information

Author notes

  1. Ke Zhang and Youhui Qiu contributed equally to this work.

Authors and Affiliations

  1. Department of Obstetrics and Gynaecology, School of Medicine, Zhongda Hospital, Southeast University, Nanjing, 210009, China

    Ke Zhang, Songwei Feng, Han Yin, Qi Liu, Yuxin Zhu & Yang Shen

  2. Institute for AI in Medicine, School of Artificial Intelligence, Nanjing University of Information Science and Technology, Nanjing, China

    Youhui Qiu, Haoyu Cui & Xiangxue Wang

  3. Department of Pathology, School of Medicine, Zhongda Hospital, Southeast University, Nanjing, China

    Xiaoying Wei & Guoqing Wang

Authors
  1. Ke Zhang
    View author publications

    You can also search for this author inPubMed Google Scholar

  2. Youhui Qiu
    View author publications

    You can also search for this author inPubMed Google Scholar

  3. Songwei Feng
    View author publications

    You can also search for this author inPubMed Google Scholar

  4. Han Yin
    View author publications

    You can also search for this author inPubMed Google Scholar

  5. Qi Liu
    View author publications

    You can also search for this author inPubMed Google Scholar

  6. Yuxin Zhu
    View author publications

    You can also search for this author inPubMed Google Scholar

  7. Haoyu Cui
    View author publications

    You can also search for this author inPubMed Google Scholar

  8. Xiaoying Wei
    View author publications

    You can also search for this author inPubMed Google Scholar

  9. Guoqing Wang
    View author publications

    You can also search for this author inPubMed Google Scholar

  10. Xiangxue Wang
    View author publications

    You can also search for this author inPubMed Google Scholar

  11. Yang Shen
    View author publications

    You can also search for this author inPubMed Google Scholar

Contributions

All authors had final approval of the submitted versions.

Corresponding authors

Correspondence to Xiangxue Wang or Yang Shen.

Ethics declarations

Ethics approval and consent to participate

The studies involving human participants were reviewed and approved by the Ethics Committees and Institutional Review Boards of Zhongda Hospital Southeast University (2024ZDSYLL268-P01). The patients provided their written informed consent to participate in this study.

Conflict of interest

There are no conflicts of interest regarding the publication of this article.

Additional information

Publisher’s note

Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.

Electronic supplementary material

Below is the link to the electronic supplementary material.

Supplementary Material 1

Rights and permissions

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/.

Reprints and permissions

About this article

Check for updates. Verify currency and authenticity via CrossMark

Cite this article

Zhang, K., Qiu, Y., Feng, S. et al. Development of model for identifying homologous recombination deficiency (HRD) status of ovarian cancer with deep learning on whole slide images. J Transl Med 23, 267 (2025). https://doiorg.publicaciones.saludcastillayleon.es/10.1186/s12967-025-06234-7

Download citation

  • Received:

  • Accepted:

  • Published:

  • DOI: https://doiorg.publicaciones.saludcastillayleon.es/10.1186/s12967-025-06234-7

Keywords

Journal of Translational Medicine

ISSN: 1479-5876

Contact us