Histopathologic basis of a deep learning pelvic computed tomography model for prognostic prediction among patients with advanced high-grade serous ovarian carcinoma

Introduction

Extensive studies have been conducted on forecasting the prognosis of patients with ovarian cancer prior to the initiation of treatment. Key prognostic indicators have been established, including International Federation of Gynecology and Obstetrics (FIGO) stage, histopathological tumor grade, human epididymis 4 (HE4) level, and carbohydrate antigen 125 (CA-125) level. As the most prevalent subtype of ovarian cancer, high-grade serous ovarian carcinoma (HGSOC) exhibits a biological behavior characterized by distinct metastatic patterns and molecular features. Its transcoelomic dissemination, propensity for TP53 mutation, a high proliferative index, and homologous recombination deficiency, collectively underpin its invasiveness and heterogeneity (1,2). With the integration of artificial intelligence (AI) technology into medical practice, radiology has demonstrated considerable value in supporting diagnosis and prognosis (3,4). The two primary AI-based imaging analysis techniques, deep learning (DL) and radiomics, enable the quantification of image-derived phenotypes across entire lesions. DL can autonomously learn complex image patterns via neural networks, while radiomics involves the conversion of images into mineable data through high-throughput feature extraction; however, radiogenomics transcends both of these approaches by establishing a bridge between imaging phenotypes and molecular histology. Its integrative framework deciphers the biological basis of images by linking specific radiological traits to genetic, protein, or structural alterations. These techniques can predict patient outcomes through use of mathematically defined image descriptors and thus exhibit a remarkable capacity to correlate intricate visual features of tissue organization with molecular disease signatures (5,6). They facilitate prognostic prediction and screening for patients with inherited cancer syndromes using clinically established and cost-effective radiological images. Multiple studies examining DL, radiomics, or textural features suggest that prognostic information may exist beyond the anatomical details obtained from computed tomography (CT) scans (7-10). A recently developed preoperative DL prediction model, leveraging contrast-enhanced CT, was introduced to predict peritoneal recurrence and disease-free survival in patients with advanced ovarian cancer (11). This DL model demonstrated satisfactory performance in predicting the outcomes of patients with advanced ovarian cancer, accurate automatic tumor segmentation, and prognostic prediction. However, its interpretability remains limited, and it cannot establish a direct link between predictions and survival outcomes.

A significant hurdle in survival prediction models is clinicians’ lack of confidence, stemming from these models’ limited interpretability. This is particularly evident in DL models, in which the underlying medical reasoning process remains opaque and bereft of a solid clinical foundation. Although efforts have been made to explain these models from a mathematical perspective, their black-box nature continues to be an obstacle to gaining clinicians’ trust and acceptance (12,13) and may hinder their integration into real-world clinical practice. Effectively interpreting these models necessitates comprehension of the intricate relationship between input, predictions, and patient outcomes. Histopathology, immunohistochemistry, and genetic information can serve as valuable tools for deciphering the connections between diverse data sets and may potentially enhance the interpretability of DL models and their clinical utility.

To overcome the interpretability deficit in DL, we established a biological rationale linking DL outputs and model-driven CT features by demonstrating the histopathological correlation between them. We evaluated six well-known prognostic histopathological manifestations of HGSOC, namely, P53, P16, Ki-67, and invasion of the omental, rectal, and pelvic wall. Our objective was to investigate and visualize the correlation between CT-driven features extracted from the DL model and histopathological risk factors by using unsupervised clustering analysis and to further determine the correlation between each feature and clinical presentation by using regression and correlation analyses. We present this article in accordance with the STROBE reporting checklist (available at https://qims.amegroups.com/article/view/10.21037/qims-2025-297/rc).

Methods

Ethical approval was obtained from the institutional review boards of Tianjin Medical University Cancer Institute and Hospital, Shanxi Tumor Hospital and Baoding No. 1 Central Hospital (approval numbers: bc2021049, TY2002008, 2023047), and individual consent for this retrospective analysis was waived. This study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments.

Patients

Patients who underwent interval debulking surgery for HGSOC (FIGO stage IIc–IV) and platinum-based neoadjuvant chemotherapy (NACT) were retrospectively enrolled from three tertiary care centers (Tianjin Medical University Cancer Institute and Hospital from January 2013 to December 2019, Shanxi Tumor Hospital from February 2014 to December 2019, and Baoding No. 1 Central Hospital from April 2015 to December 2019). The exclusion criteria were low-grade serous ovarian cancer and other rarer pathological types of ovarian cancer, FIGO stage I–IIb, absence of an available and assessable preoperative CT and CT obtained after needle biopsy, a CT and NACT interval of more than 30 days, and malignant neoplasms other than ovarian cancer. The final sample comprised 418 women, which was overlapped with the patient sample of a preliminary multicenter two-dimensional DL study on the prediction of recurrence and prognosis. Figure 1 shows the overall study design.

Figure 1 Flowchart of patient inclusion. HGSOC, high-grade serous ovarian cancer; LGSOC, low-grade serous ovarian cancer.

Data collection

Figure 2 portrays a comprehensive overview of our interpretability analysis. Our approach consisted three steps: data acquisition, modeling, and interpretability analysis. Histopathologic analysis was conducted by gynecologic pathologists, and histopathologic reports were obtained to determine the immunohistochemical expression and detailed histologic findings. Specifically, this included the extent of tumor invasion observed intraoperatively; the presence of lymphatic invasion; and the status of human epidermal growth factor receptor 2 (HER2), estrogen receptor (ER), progesterone receptor (PR), and Ki-67. Tumors were classified as ER/PR-positive with >10% immunostaining of cells, HER2-positive with 3+ hematoxylin-eosin staining, and Ki-67-positive with ≥14% staining of cells.

Figure 2 Schematic overview of the interpretability analysis. CT, computed tomography; FC, fully connected layer; PC, principal component; PCA, principal component analysis.

The following semantic CT features were also obtained: the FIGO stage descriptor based on the according to the 2018 edition staging system for ovarian cancer; the radiologic nodule type according to solid components presence (i.e., solid, subsolid, and nonsolid); and presence of lymph node enlargement, thoracoabdominal fluid, and other organ invasion or metastasis. Preliminary information, including age, HE4 level, and CA-125 level, was also collected.

CT image acquisition and tumor annotation

CT scans were performed with the Discovery CT750 HD (GE HealthCare, Chicago, IL, USA), Lightspeed 16 (GE HealthCare), or SOMATOM Definition AS (Siemens Healthineers, Erlangen, Germany) CT scanners. Contrast-enhanced CT was performed following intravenous contrast administration, with the arterial and portal venous phases initiated at delays of 25 and 60 s, respectively. CT (abdomen and pelvis) was performed with patients in the supine position and covering the lung base to the groin. The original cross-sectional CT images of the non-contrast-enhanced and arterial and portal phases were used for analyses. For patients with multiple adnexal masses, the dominant mass, considered to be the lesion with the most complex morphology or the largest size based on CT findings, was chosen as the subject of evaluation. Tumor annotations were performed on each scan via commercially available software (3D Slicer software version 4.10.2, www.Slicer.org). Three plane patches were then obtained for each tumor and were used as the input for the DL model. Detailed information on the CT image acquisition and tumor annotation is provided in Appendix 1.

DL prediction model of recurrence and prognosis

DL models were trained with preoperative pelvic CT from 515 patients to determine whether recurrence would occur in patients after definitive treatment and time and site of recurrence. The output was the probability of recurrence and peritoneal recurrence and timing of relapse. Consequently, the model was applied to HGSOC with an output equal to the cumulative risk of adverse events after the procedure, which ranged from 0% to 100%. The DL model employed included a multitask convolutional neural network (CNN) consisting of autosegmentation and autoclassification pipeline, with patches containing tumor regions as input. In this study, the model was used for inference only; none of the training procedures, including fine-tuning or transfer learning, were performed. Detailed information on the model can be found elsewhere (12). Other information on the model, including the inference process, is provided in Appendix 2.

Unsupervised clustering with CT features extracted by the DL model

To determine whether DL model-driven CT features can be used to identify a subset of patients with histopathological risk factors, an unsupervised clustering analysis was performed. A total of 1,280 high-dimensional CT imaging features were extracted from the fully connected layer of the DL model. Next, feature dimension reduction was performed to reduce the overfitting and bias. First, the Mann-Whitney test was used to select features that were highly correlated with the biomarkers, with P<0.05 serving as the significance threshold. Agglomerative hierarchical clustering, a bottom-up and data-driven approach, was subsequently applied. According to a hierarchical clustering dendrogram, the number of clusters was determined to be four. For two-dimensional visualization, principal component analysis (PCA) was performed to reduce uncertainty in the DL model-driven CT features. Moreover, we believe that DL features have variable importance across different pathologies and immunohistochemical states, and thus a feature selection step for each indicator was performed.

Evaluation of the model in predicting overall survival (OS)

OS was defined as the time to death from any cause. Patients who were followed up for more than 5 years or who died within 5 years were included in the study. Ultimately, 344 patients with HGSOC who completed follow-up were deemed eligible for inclusion. The median of the PCA score and the four groups of unsupervised cluster analysis were selected as the criteria for grouping the patients with HGSOC. The OS of patients with HGSOC was also analyzed according to FIGO staging.

Statistical analysis

Continuous variables were compared via the t test, while categorical variables were analyzed with the χ² test or Fisher’s exact test. To assess the relationship between the unsupervised group and histopathological risk factors, univariate logistic regression analysis was applied. Subsequently, to determine the histopathological correlates of the added value of the DL model output and FIGO staging, multivariate logistic regression analysis was performed with each histopathological risk factor as the dependent variable and the model output and FIGO staging as covariates. Survival curves, generated by the Kaplan-Meier method, were compared via the log-rank test. All statistical analyses were conducted with R version 3.4.0 (The R Foundation of Statistical Computing, Vienna, Austria) and MedCalc version 20.1 (MedCalc Software, Ostend, Belgium), with a two-tailed P value <0.05 being considered statistically significant.

Results

Patient characteristics and surgical histopathologic findings

From January 2013 to December 2019, 81 (19.1%) of 423 patients in the primary cohort and 16 (17.4%) of 92 patients in the validation cohort were excluded due to being diagnosed with ovarian clear cell adenocarcinoma (n=37), low-grade serous ovarian carcinoma (n=6), endometrioid carcinoma (n=29), or other rare pathological types (n=25). Finally, 418 patients were included in the study (Figure 1).

Among the 418 patients [median age 55 years, interquartile range (IQR) 30–77 years], 44.0% (184 of 418) were solid tumors, 53.8% (225 of 418) were subsolid tumors, and 2.2% (9 of 418) were nonsolid tumors. On preoperative CT, 48.6% (203 of 418) had invasion of the omentum and 41.6% (174 of 418) had lymph node invasion. In terms of FIGO stage, 25.1% (105 of 418 patients) were FIGO stage IIc, 37.3% (156 of 418 patients) were stage III, and 36.7% (157 of 418 patients) were stage IV. According to surgical histopathologic findings, 53.8% (225 of 418) had invasion to the omentum, 51% (213 of 418) had invasion to the rectum, 47.1% (197 of 418) had bladder invasion, and 46.2% (193 of 418) had pelvic wall effusion. Detailed information, including the proportions of each histopathologic and immunohistochemical risk factor, is provided in Tables 1,2.

Unsupervised clustering and associations with histopathologic and immunohistochemical risk factors

Patients were stratified into four distinct clusters (clusters 1–4) through hierarchical agglomerative clustering (Figure S1 and Table S1). Of the four clusters, unsupervised clusters 3 and 4 were associated with six histopathologic and immunohistochemical risk factors: positive status of P53, P16, and Ki-67 and invasion of the omentum, rectum, and pelvic wall [P<0.05; β, odds ratios (ORs), and Nagelkerke pseudo R² are presented in Table 3]. Thus, the DL model was able to identify a subset of patients among those with HGSOC with high-risk histopathological presentation and highly aggressive activity.

In the principal component biplots, patients from cluster 3 were primarily distributed in the lower-right quadrant, while those in cluster 4 were located in the upper- and lower-left quadrants (Figure 3). A similar distribution was observed for the status of P53, P16, and Ki-67, along with omentum invasion, rectum invasion, and pelvic wall effusion (Figure 4).

Figure 3 Principal component biplots (X-axis, principal component 1; Y-axis, principal component 2) of 1,280 deep learning model-driven CT features. The data points of patients with (A) HGSOC and (B) unsupervised clusters. CT, computed tomography; HGSOC, high-grade serous ovarian cancer.

Figure 4 Principal component biplots (X-axis, principal component 1; Y-axis, principal component 2) of the 1,280 deep learning model-driven CT features. The data points of patients with (A) Ki-67 expression, (B) P16 expression, (C) P53 expression, (D) omentum invasion, (E) rectum invasion, and (F) pelvic wall invasion. CT, computed tomography.

DL model output and its independent associations with histopathologic and immunohistochemical risk factors

In the multivariate logistic regression analysis with adjustment for the FIGO stage, the DL model output was independently associated with P53 status [OR for 1% increment of the model output, 1.9642; 95% confidence interval (CI): 1.2412–3.1082; P=0.0039], P16 status (OR 2.3446; 95% CI: 1.5445–3.5592; P=0.0001), Ki-67 level (OR 10.0433; 95% CI: 5.3525–18.8450; P<0.0001), invasion to the omentum (OR 2.5995; 95% CI: 1.7175–3.9342; P<0.0001), invasion to the rectum (OR 2.3568; 95% CI: 1.5614–3.5574; P<0.0001), and pelvic wall effusion (OR 2.0779; 95% CI: 1.3769–3.1360; P=0.0005) (Table 4). The detailed results are provided in Tables S2-S7. In univariate and multivariate logistic regression analyses without adjustments for the FIGO stage, the DL model output was independently associated with the six histopathologic and immunohistochemical risk factors (P<0.01). In the univariate analyses, however, the model output did not demonstrate added value over the CT features in explaining P53 status (OR 1.1508; 95% CI: 0.690–1.918; P=0.590), P16 status (OR 1.4246; 95% CI: 0.900–2.256; P=0.131), invasion to the rectum (OR 1.4878; 95% CI: 0.920–2.406; P=0.1051), or pelvic wall effusion (OR 1.4445; 95% CI: 0.896–2.3279; P=0.1309) (Table S8). Figure 5 includes a box plot showing the DL output stratified by histopathological and immunohistochemical risk factors.

Figure 5 The histopathologic correlates with DL output. (A) Ki-67 expression, (B) P16 expression, (C) P53 expression, (D) omentum invasion, (E) rectum invasion, and (F) pelvic wall invasion. DL, deep learning.

DL model features and the independent associations with histopathologic and immunohistochemical risk factors

With the feature dimension reduction step, 1,140 DL features were excluded from further analysis because their correlation coefficients were above the predefined 0.05 threshold. Thus, there were 140 features remaining from the 1,280 (10.9%) originally extracted features after this first feature selection step. Before PCA was conducted, the top five most discriminative DL features as ranked according to the Spearman correlation coefficient for each histopathologic and immunohistochemical risk factor, were identified and are reported in Table S9. Among the 30 top 5 DL features, 17 were duplicates, which confirmed that the DL feature selection process could detect important features. The relationship between histopathologic and immunohistochemical risk factors and DL features is shown in Figure 6A. Analysis of the Spearman correlation network revealed an absence of strong collinearity among the selected features, with all correlation coefficients falling within the moderate range between −0.6 and 0.6 (Figure 6B and Figure S2).

Figure 6 Feature correlation analysis. (A) The Spearman correlation coefficient heat maps show the relationships between each histopathologic risk factors and the deep learning features. (B) The Spearman correlation coefficient network diagrams show the relationships between selected features in HGSOC. All the imaging-derived features were independent from clinical characteristics as no correlation coefficient >0.8 was observed. Each feature was an independent predictor as there was no correlation coefficient >0.8. HGSOC, high-grade serous ovarian cancer.

Prognostic value of PCA and unsupervised clustering

Overall, the distribution of the patients with HGSOC with low and high PCA scores was 50.0% (n=172) and 50.0% (n=172), respectively. A lower PCA score was associated with worse OS (OR =1.7842; 95% CI: 1.3703–2.3232; P<0.0001). Patients with a lower PCA score had a significantly worse OS (median OS 32.618 months, IQR 29.700–35.537 months) compared to those with high PCA score (median OS 43.246 months, IQR 40.558–45.933 months), and the former were placed in the high-risk patient subgroup. Similarly, with unsupervised clustering groupings used as the criterion, patients in unsupervised cluster 4 all showed a shorter OS (all P values <0.001). Notably, because the number of patients in the unsupervised clusters 1 and 2 was small and given the possibility of randomness and bias in the survival analysis of these two groups, they were combined. After this adjustment, it was found that the low-risk patients had better OS (median OS 43.645 months, IQR 37.226–50.064 months) compared to unsupervised clusters 3 and 4 (all P values <0.001). Unsupervised cluster 4 was associated with the worst OS (median OS 32.198 months, IQR 29.167–35.228 months) (Figure 7). A similar survival pattern was observed in the other FIGO stages (all P values <0.001) (Figure S3).

Figure 7 Kaplan-Meier analyses of overall survival with (A) PCA and unsupervised clusters (B) being used as the grouping criterion. PCA, principal component analysis.

Discussion

In this study, we investigated whether preoperative CT-based DL prediction models could provide imaging alternatives to histopathological risk factors among patients with HGSOC in a retrospective surgical sample of 418 patients. Our data confirmed the value of DL analysis in assessing patients with HGSOC and further enriches the body of evidence related to DL-based approaches in gynecological cancer. When unsupervised clustering of patients was performed with the 1,280 DL model-driven CT features, unsupervised clusters 3 and 4 were correlated with histopathological risk factors and showed similar distributions to patients with histopathology risk factors in the PCA plot. In the multivariate logistic regression analyses, the DL model output was independently associated with P53 status, P16 status, and Ki-67 level, along with invasion to the omentum, rectum, and pelvic wall (all P values <0.05). In the survival analysis, we found that a lower PCA score and unsupervised cluster 4 were associated with a worse OS.

Unlike the classic and well-studied pattern of hematogenous metastasis common to most cancers, the biological behavior of ovarian cancer is unique. For HGSOC, once cancer cells are isolated from the primary ovarian tumor as single cells or clusters, it is believed that they metastasize by a passive mechanism, moving to the peritoneum and omentum through the physiological movement of peritoneal fluid (14-16). The aggressiveness of HGSOC is determined by the progression of peritoneal metastases, and most patients have a poor and an incurable condition (17). The omentum contains immune aggregates, embedded in adipose tissue, that support tumor growth and promote tumor aggressiveness, proliferation, and resistance to cancer therapy (18,19). In addition, tumor invasion of the rectum is a well-known indicator of a poor prognosis (20). Given this pathologic evidence, the output of the DL model represents high-risk histopathologic features that appear to preoperatively predict poor outcomes in patients with HGSOC. In the cluster analysis, unsupervised clusters 3 and 4 were associated with pathologic risk factors, and patients who developed metastases were separated on the PCA plot. In addition, immunohistochemical analysis confirmed that patients with a high expression of P53, P16, and Ki-67 had a greater malignancy of disease, which is consistent with the findings of Němejcová et al. (21). Thus, we propose that our model can be used to preoperatively stratify patients according to risk, identify candidates for neoadjuvant therapy or cytoreductive surgery, and guide surgical planning.

Importantly, the model outputs were independent and complementary to clinical FIGO stage, suggesting that our DL-based CT features can yield additional information that would be difficult to obtain clinically in the risk stratification of patients with HGSOC. Therefore, we believe that our model can effectively predict the survival and prognosis of patients with HGSOC and optimize postoperative detection and management (for example, closer monitoring or switching to other treatment strategies for high-risk patients).

In the unsupervised cluster regression analyses, the expression of ER, PR and HER2 was not associated with the unsupervised cluster clustering, similar to the histopathological risk factors. A possible explanation is that ER and PR positivity is a reflection of the disease course being governed by estrogenic progesterone, which in ovarian cancer tends to guide the pathological type of tumor and the patient’s treatment regimen but is of relatively limited value in determining the prognosis (22,23). However, as only patients with HGSOC were included in this study, there might have been bias in regards to in ER, PR, and HER2 expression.

Deep neural networks are often regarded as “black-box” models due to their nonlinear characteristics and the difficulty in explaining their internal working mechanism (24,25). The limitation mainly involves the lack of mathematical tools to diagnose and evaluate the feature expressiveness of the network, and thus it remains challenging to explain the information-processing characteristics of the different currently available neural network models. However, for risk prediction, especially in the medical field, domain-specific explanations are necessary. DL, as one of the key technologies involved in big data analytics, has achieved valuable practical results in a number of fields, including speech recognition, visual object recognition, target detection, and drug discovery, among others, by virtue of its ability to overcome the limitations of traditional machine learning algorithms that rely on the establishment and screening of human features (26-29). In order to advance the utility of DL in the medical field and promote its use in clinical work, providing an explanation related to the pathophysiology of the disease is necessary. The primary significance of our study lies in the establishment of a tangible link between DL-driven CT biomarkers and the model’s decision-making process, thereby enabling domain-specific interpretation of its otherwise opaque predictions. We found that DL features representing patient prognosis were significantly associated with histopathological risk factors. This finding suggests that our model can successfully extract prognostic information from pelvic CT images—a capability that enhances the model’s credibility while contributing to more informed, personalized treatment strategies.

However, our study involved certain limitations which should be acknowledged. First, although the sample used to develop the DL model was derived from multiple institutions, the retrospective nature of the dataset limits the generalizability of our results. Second, we attempted to assess the correlation between the information in the fully connected layer and the output layer, but this approach ignored the potentially more valid information contained in other layers of the neural network. Third, we only included patients with ovarian cancer of a specific pathologic type, and further evaluation of all ovarian cancer types is needed in future work. Finally, we did not perform a priori sample size calculations.

Conclusions

DL model-driven CT features based on preoperative CT images for predicting the prognosis of patients with HGSOC showed a strong correlation with histopathology, suggesting that our DL model can effectively identify patients with high-risk HGSOC.

Acknowledgments

None.

Footnote

Reporting Checklist: The authors have completed the STROBE reporting checklist. Available at https://qims.amegroups.com/article/view/10.21037/qims-2025-297/rc

Data Sharing Statement: Available at https://qims.amegroups.com/article/view/10.21037/qims-2025-297/dss

Funding: The present study was sponsored by Natural Science Foundation of China (Nos. 81801781, 82072004), Tianjin Key Medical Discipline (Specialty) Construction Project (Nos. TJYXZDXK-009A, TJYXZDXK-012A).

Conflicts of Interest: All authors have completed the ICMJE uniform disclosure form (available at https://qims.amegroups.com/article/view/10.21037/qims-2025-297/coif). The authors have no conflicts of interest to declare.

Ethical Statement: The authors are accountable for all aspects of the work in ensuring that questions related to the accuracy or integrity of any part of the work are appropriately investigated and resolved. Ethical approval was obtained from the institutional review boards of Tianjin Medical University Cancer Institute and Hospital, Shanxi Tumor Hospital and Baoding No. 1 Central Hospital (approval numbers: bc2021049, TY2002008, 2023047), and individual consent for this retrospective analysis was waived. This study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments.

Open Access Statement: This is an Open Access article distributed in accordance with the Creative Commons Attribution-NonCommercial-NoDerivs 4.0 International License (CC BY-NC-ND 4.0), which permits the non-commercial replication and distribution of the article with the strict proviso that no changes or edits are made and the original work is properly cited (including links to both the formal publication through the relevant DOI and the license). See: https://creativecommons.org/licenses/by-nc-nd/4.0/.

References

1. Horn LC, Mayr D, Brambs CE, Einenkel J, Sändig I, Schierle K. Grading of gynecological tumors: Current aspects. Pathologe 2016;37:337-51. [Crossref] [PubMed]
2. Van Gorp T, Cadron I, Despierre E, Daemen A, Leunen K, Amant F, Timmerman D, De Moor B, Vergote I. HE4 and CA125 as a diagnostic test in ovarian cancer: prospective validation of the Risk of Ovarian Malignancy Algorithm. Br J Cancer 2011;104:863-70. [Crossref] [PubMed]
3. Chen H, Yang BW, Qian L, Meng YS, Bai XH, Hong XW, He X, Jiang MJ, Yuan F, Du QW, Feng WW. Deep Learning Prediction of Ovarian Malignancy at US Compared with O-RADS and Expert Assessment. Radiology 2022;304:106-13. [Crossref] [PubMed]
4. Wang CW, Chang CC, Lee YC, Lin YJ, Lo SC, Hsu PC, Liou YA, Wang CH, Chao TK. Weakly supervised deep learning for prediction of treatment effectiveness on ovarian cancer from histopathology images. Comput Med Imaging Graph 2022;99:102093. [Crossref] [PubMed]
5. Wang S, Liu Z, Rong Y, Zhou B, Bai Y, Wei W, Wei W, Wang M, Guo Y, Tian J. Deep learning provides a new computed tomography-based prognostic biomarker for recurrence prediction in high-grade serous ovarian cancer. Radiother Oncol 2019;132:171-7. [Crossref] [PubMed]
6. Zhang Z, Wei Z, Zhao L, Gu C, Meng Y. Assessing the clinical utility of multi-omics data for predicting serous ovarian cancer prognosis. J Obstet Gynaecol 2023;43:2171778. [Crossref] [PubMed]
7. Wang R, Dai W, Gong J, Huang M, Hu T, Li H, Lin K, Tan C, Hu H, Tong T, Cai G. Development of a novel combined nomogram model integrating deep learning-pathomics, radiomics and immunoscore to predict postoperative outcome of colorectal cancer lung metastasis patients. J Hematol Oncol 2022;15:11. [Crossref] [PubMed]
8. Skrede OJ, De Raedt S, Kleppe A, Hveem TS, Liestøl K, Maddison J, Askautrud HA, Pradhan M, Nesheim JA, Albregtsen F, Farstad IN, Domingo E, Church DN, Nesbakken A, Shepherd NA, Tomlinson I, Kerr R, Novelli M, Kerr DJ, Danielsen HE. Deep learning for prediction of colorectal cancer outcome: a discovery and validation study. Lancet 2020;395:350-60. [Crossref] [PubMed]
9. Huang YQ, Liang CH, He L, Tian J, Liang CS, Chen X, Ma ZL, Liu ZY. Development and Validation of a Radiomics Nomogram for Preoperative Prediction of Lymph Node Metastasis in Colorectal Cancer. J Clin Oncol 2016;34:2157-64. [Crossref] [PubMed]
10. Boehm KM, Aherne EA, Ellenson L, Nikolovski I, Alghamdi M, Vázquez-García I, et al. Multimodal data integration using machine learning improves risk stratification of high-grade serous ovarian cancer. Nat Cancer 2022;3:723-33. [Crossref] [PubMed]
11. Yin R, Dou Z, Wang Y, Zhang Q, Guo Y, Wang Y, Chen Y, Zhang C, Li H, Jian X, Qi L, Ma W. Preoperative CECT-Based Multitask Model Predicts Peritoneal Recurrence and Disease-Free Survival in Advanced Ovarian Cancer: A Multicenter Study. Acad Radiol 2024;31:4488-98. [Crossref] [PubMed]
12. van der Velden BHM, Kuijf HJ, Gilhuijs KGA, Viergever MA. Explainable artificial intelligence (XAI) in deep learning-based medical image analysis. Med Image Anal 2022;79:102470. [Crossref] [PubMed]
13. Lee M, Sanz LRD, Barra A, Wolff A, Nieminen JO, Boly M, Rosanova M, Casarotto S, Bodart O, Annen J, Thibaut A, Panda R, Bonhomme V, Massimini M, Tononi G, Laureys S, Gosseries O, Lee SW. Quantifying arousal and awareness in altered states of consciousness using interpretable deep learning. Nat Commun 2022;13:1064. [Crossref] [PubMed]
14. Lengyel E. Ovarian cancer development and metastasis. Am J Pathol 2010;177:1053-64. [Crossref] [PubMed]
15. Bella Á, Arrizabalaga L, Di Trani CA, Fernández-Sendin M, Teijeira A, Russo-Cabrera JS, Melero I, Berraondo P, Aranda F. Omentum: Friend or foe in ovarian cancer immunotherapy? Int Rev Cell Mol Biol 2022;371:117-31. [Crossref] [PubMed]
16. Bose S, Saha P, Chatterjee B, Srivastava AK. Chemokines driven ovarian cancer progression, metastasis and chemoresistance: Potential pharmacological targets for cancer therapy. Semin Cancer Biol 2022;86:568-79. [Crossref] [PubMed]
17. Nieman KM, Kenny HA, Penicka CV, Ladanyi A, Buell-Gutbrod R, Zillhardt MR, Romero IL, Carey MS, Mills GB, Hotamisligil GS, Yamada SD, Peter ME, Gwin K, Lengyel E. Adipocytes promote ovarian cancer metastasis and provide energy for rapid tumor growth. Nat Med 2011;17:1498-503. [Crossref] [PubMed]
18. Mukherjee A, Chiang CY, Daifotis HA, Nieman KM, Fahrmann JF, Lastra RR, Romero IL, Fiehn O, Lengyel E. Adipocyte-Induced FABP4 Expression in Ovarian Cancer Cells Promotes Metastasis and Mediates Carboplatin Resistance. Cancer Res 2020;80:1748-61. [Crossref] [PubMed]
19. Ladanyi A, Mukherjee A, Kenny HA, Johnson A, Mitra AK, Sundaresan S, Nieman KM, Pascual G, Benitah SA, Montag A, Yamada SD, Abumrad NA, Lengyel E. Adipocyte-induced CD36 expression drives ovarian cancer progression and metastasis. Oncogene 2018;37:2285-301. [Crossref] [PubMed]
20. O'Hanlan KA, Kargas S, Schreiber M, Burrs D, Mallipeddi P, Longacre T, Hendrickson M. Ovarian carcinoma metastases to gastrointestinal tract appear to spread like colon carcinoma: implications for surgical resection. Gynecol Oncol 1995;59:200-6. [Crossref] [PubMed]
21. Němejcová K, Šafanda A, Bártů MK, Michálková R, Drozenová J, Fabian P, Hausnerová J, Laco J, Matěj R, Méhes G, Škapa P, Stružinská I, Dundr P. A comprehensive immunohistochemical analysis of 26 markers in 250 cases of serous ovarian tumors. Diagn Pathol 2023;18:32. [Crossref] [PubMed]
22. Vetter M, Stadlmann S, Bischof E, Georgescu Margarint EL, Schötzau A, Singer G, Heinzelmann-Schwarz V, Montavon C. Hormone Receptor Expression in Primary and Recurrent High-Grade Serous Ovarian Cancer and Its Implications in Early Maintenance Treatment. Int J Mol Sci 2022;23:14242. [Crossref] [PubMed]
23. Ji R, Li Y, He C, Zhu X, He A, Lu Y. Detection and analysis of multiple biomarkers in ovarian cancer: clinical significance in diagnosis, treatment, and prognosis evaluation. Gland Surg 2020;9:2175-86. [Crossref] [PubMed]
24. Shao Y, Cheng Y, Shah RU, Weir CR, Bray BE, Zeng-Treitler Q. Shedding Light on the Black Box: Explaining Deep Neural Network Prediction of Clinical Outcomes. J Med Syst 2021;45:5. [Crossref] [PubMed]
25. Cui S, Tseng HH, Pakela J, Ten Haken RK, El Naqa I. Introduction to machine and deep learning for medical physicists. Med Phys 2020;47:e127-47. [Crossref] [PubMed]
26. Ting DSW, Peng L, Varadarajan AV, Keane PA, Burlina PM, Chiang MF, Schmetterer L, Pasquale LR, Bressler NM, Webster DR, Abramoff M, Wong TY. Deep learning in ophthalmology: The technical and clinical considerations. Prog Retin Eye Res 2019;72:100759. [Crossref] [PubMed]
27. Wang J, Sun K, Cheng T, Jiang B, Deng C, Zhao Y, Liu D, Mu Y, Tan M, Wang X, Liu W, Xiao B. Deep High-Resolution Representation Learning for Visual Recognition. IEEE Trans Pattern Anal Mach Intell 2021;43:3349-64. [Crossref] [PubMed]
28. Wu L, Shang R, Sharma P, Zhou W, Liu J, Yao L, et al. Effect of a deep learning-based system on the miss rate of gastric neoplasms during upper gastrointestinal endoscopy: a single-centre, tandem, randomised controlled trial. Lancet Gastroenterol Hepatol 2021;6:700-8. [Crossref] [PubMed]
29. Zhong F, Xing J, Li X, Liu X, Fu Z, Xiong Z, Lu D, Wu X, Zhao J, Tan X, Li F, Luo X, Li Z, Chen K, Zheng M, Jiang H. Artificial intelligence in drug design. Sci China Life Sci 2018;61:1191-204. [Crossref] [PubMed]