In the gynecological field, with an estimated 184,799 deaths worldwide annually, ovarian cancer is one of the most common and deadliest tumors. Among them, high-grade serous carcinoma subtype (HGSOC) is the most aggressive form and accounts for the majority of mortality (1). Despite advances in HGSOC therapy, such as surgery, chemotherapy, targeted therapy and immunotherapy, the 5-year overall survival is still substantial, less than 50% (2). Currently, stratification of HGSOC risk is still based on the stage of International Federation of Gynecology and Obstetrics (FIGO) (3), but because of the spatial and temporal heterogeneity of tumors, clinical biomarkers provide only partial information (4).

Currently, with its capacity to visualize a cancer’s appearance at a macroscopic level noninvasively, medical imaging contains more prognostic information for the primary tumor (5). Growing evidence suggests that computed tomography (CT) contains more mineable high-dimensional data which could promote personalized care and survival prediction in cancer patients (6–8). Despite the advantages that CT could quantify tumor shape and texture information, the existing hand-crafted feature engineering is difficult to extract the full intrinsic characteristics and prone to human biases (4, 8).

In recent years, with the development of optimization techniques and the improvement in computing devices, deep learning is becoming a popular method in medical image analysis (9). Because of its unique ability to learn features from raw data, the intrinsic characteristics of images have been mined automatically, thus reducing information redundancy and aiding clinical decision (10, 11). With the Transformer architecture, the Vision Transformer (ViT) has been shown to model long-range dependency among pixels and demonstrated the state-of-the-art (SOTA) performance in image classification task (12, 13). The attention mechanism in ViT model allows integration of global information but not of the local information from the CT images, which could have a significant impact on classification tasks. Existing literature describes the use of the ViT model to predict medical targets, such as emphysema classification (14) and COVID-19 diagnosis (15).

In this study, we developed a Vit-based deep learning model to predicting overall survival in patients with HGSOC based on preoperative CT images. Instead of delineating precise tumor boundaries which often used in conventional radiomics methods, our model requires only a rectangle region of tumor, thus reducing the interobserver error and manual segmentation time. Moreover, integrating the prediction features learned from CT images, clinicopathological and hematological markers, we established a comprehensive nomogram aiming to provide a non-invasive individualized recurrence prediction model in HGSOC.

Our work proposed a novel and innovative deep learning framework for prognosis and prediction in HGSOC based on preoperative CT images. Combined Vision Transformer with RNNs, our model extracted an image score from preoperative CT images for each patient, which was then shown to be an independent prognostic factor for HGSOC (AUC is 0.822 in the training cohort and 0.823 in the validation cohort) and had prognostic value above other clinical characters. Then, incorporated clinical characters and model-based image score, a concise nomogram was performed for survival prediction and intended to serve as a practical guide to clinicians when recommending an appropriate management strategy for HGSOC patients.

In recent years, radiomics has grown rapidly as a medical research field due to the combination of radiographic images and data about clinical outcomes. In earlier radiomics studies, medical images are assessed visually by trained physicians for the purpose of detecting, characterizing, and monitoring diseases, which often relies on education and experience and can be subjective at times. Deep learning, however, excels at identifying complex patterns and can provide automated quantitative assessments in contrast to qualitative reasoning (16). More meaningfully, the feature representations automatically learned by deep learning techniques hinting at the substantial clinical relevance of many of these radiographic features (17). Preoperative CT examination is very necessary for patients with solid tumors and it may contain multiple anatomic and nonanatomic elements which may enhance the capacity to prognosticate (18, 19). In ovarian cancer, preoperative CT is an important means of staging and treatment decision (20, 21), however, the prognostic information provided by manual observation is limited, we believe that through deep learning, more information about prognosis could be directly extracted and could constitute the prognostic index for patients with ovarian cancer. Several previous studies reported the efficacy of deep learning for the survival prediction of cancer on radiology images (7, 22–24), but most of them have used traditional CNNs. Wang et al. constructed a non-invasive recurrence prediction model based on CNN (8). Avesani et al. used a CNN as feature extractor to predict progression free survival and BRCA mutational status (25). Comprising 44,732 slides from 15,187 patients, Gabriele et al. developed a deep learning frame-work that combined CNNs with RNNs to diagnose prostate cancer, the semantically rich tile-level feature representations resulted from CNNs were then used in a RNN to integrate the information across the whole slide and report the final classification result (26). The AUC of this model was above 0.98 and its clinical application would allow pathologists to exclude 65–75% of slides while retaining 100% sensitivity. Although in the field of medical image analysis, CNNs have been widely adopted, they have inherent limitations. CNN is good at focusing on extraction of local information, but this means the receptive field is limited and the global feature is hard to be captured.

In contrast, mainly based on self-attention mechanisms, the primary advantage of Transformer is its global receptive field and focus on the aggregation of global information. In the past few years, Transformer have been dominant in the natural language processing field and have been used in speech recognition (27), machine translation (28), and language modeling (29). More recently, to overcome these limitations of CNN in computer vision problems, equipped with the Transformer architecture, Vision Transformer (ViT) was proposed to model long-range dependency among pixels through the self-attention mechanism (12), and has been demonstrated the state-of-the-art (SOTA) performance in a variety of vision tasks including object detection (30), classification (13), segmentation (31), and so on. At present, in the field of cancer, there are several researches using ViT for classify tasks (32–34) and cancer region detection and segmentation tasks (35). As far as we know, our study is the first to attempt to apply ViT for the survival prediction of HGSOC. Our proposed ViT-based model integrates the advantages of ViT and RNN, enabling the model to have overwhelming effect on survival prediction.

Although our model performed well, there are also several limitations. First, the CT images to construct the model were collected from only one manufacture, different CT scanners may lead to distinct image features. Second, since we still require manual tumor annotation (although only bounding box) on CT images, our model is not fully automated. Third, we lacked an additional cohort for external validation. Finally, we only focus on high-grade serous carcinoma subtype, further validation studies are warranted for other ovarian cancer subtypes. Hence, in future, a more general and more robust model which was trained by more data should be considered. And our research team is currently collecting more data from other medical centers to further validate and improve the current model. In addition to preoperative CT images and clinical information, whole slide images (WSIs) of hematoxylin and eosin-stained postoperative pathological slides were also collected to develop the multimodal model.

In conclusion, this study shows a new deep learning model which could output an independent prognostic risk score for predicting survival in patients with HGSOC through their preoperative CT images. Combined with other clinical characters, the score was used to construct a simple, yet not trivial nomogram, which would have potential as a useful tool in creating optimal individualized therapeutic approaches for HGSOC patients. Our study could help predicting the survival prognostication for HGSOC patients and may facilitate clinical decision making in the era of individualized and precision medicine.

The original contributions presented in the study are included in the article/supplementary material. Further inquiries can be directed to the corresponding authors.

The studies involving human participants were reviewed and approved by Institutional Ethics committee of Qilu Hospital. The patients/participants provided their written informed consent to participate in this study. Written informed consent was obtained from the individual(s) for the publication of any potentially identifiable images or data included in this article.

YZ: Conceptualization, methodology, and writing (original draft). FW: Conceptualization, image processing and methodology. WZ: Formal analysis, data curation and writing (review and editing). YL: Formal analysis and data curation. BY: Image processing and formal analysis. XY: Conceptualization, and supervision. TD: Conceptualization, project administration and funding acquisition. All authors contributed to the article and approved the submitted version.

This study was funded by Innovation and Development Joint Funds of Natural Science Foundation of Shandong Province (ZR2021LZL009).

The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.