Gastric cancer (GC) is one of the most common and deadly carcinomas in the world (1). Peritoneal metastasis (PM), one of the most common forms of metastasis in GC, occurs in ~53%-80% of GC patients with distant metastasis (2, 3) and is generally regarded as an incurable condition with a poor prognosis (4, 5). It was reported that the median survival time among PM patients is 3–6 months (5), and the treatment options are limited (6). New treatment strategies such as intraperitoneal chemotherapy and extensive intraoperative peritoneal lavage are associated with an improved prognosis for those patients (7–9). Therefore, noninvasive preoperative detection of PM of advanced gastric cancer (AGC) is crucial for avoiding unnecessary surgery and selecting optimal therapy in clinical practice.

CT is recommended as the first-line imaging modality for the detection of PM (10). However, PMs (<5 mm) are frequently missed on CT images (11). In 10 to 30% of patients with negative CT images, intraperitoneal metastases are found during either laparoscopy or surgical exploration (11, 12), called occult peritoneal metastases (OPMs).

Artificial intelligence, especially deep learning, has been explored in predicting PM in GC (13, 14), such as previous research from our team (14). However, the interpretation is under elucidated due to its nature of “black box”. Radiomics is an emerging field focusing on the high-dimensional mineable feature set captured from imaging data using a series of quantitative characteristic algorithms. The effective extraction and modeling of digital information is expected to aid in the assessment and differential diagnosis of gastric tumors (15–20). For OPMs in AGC, some studies (21–23) have utilized radiomics analysis of preoperative CT texture features to make a preoperative, noninvasive OPM diagnosis. All regions of interest (ROIs) were drawn manually in these studies. However, the manual annotation of many medical images is time consuming, expensive, and inefficient. In addition, most medical image annotations require a certain level of expertise, which can add additional work for the physician. Recently, we applied another annotation method—an easy-to-use, time-saving and inexpensive method referred to as the bounding box (BBOX)—that has been used in some clinical areas (24–28), such as segmentation, diagnosis, and classification, but has not yet been applied to the radiomics analysis of OPM in GC patients.

The BBOX, which contains both the tumor tissue and peritumoral area (PERI), covers more regions than manual annotation of the tumor alone. A few studies have revealed that the combined radiomic signature of intratumoral and peritumoral regions can provide valuable information for the prediction of the Lauren classification of GC (29), the pathological complete response to neoadjuvant chemoradiation of esophageal squamous cell carcinoma (30), and the outcomes and chemotherapy response in GC (31, 32).

Therefore, in this study, we hypothesized that ROI annotation by BBOX might provide adequate information for the diagnosis of OPM. We aimed to develop and validate a BBOX-based radiomics model for the preoperative, noninvasive diagnosis of OPM in AGC patients.

The radiomics processing flowchart of this study is shown in Figure 1 .

In this multicenter study, we developed a radiomics model based on 2D images annotated according to the BBOX containing the primary tumor and PERI to identify OPM in AGC patients prior to surgical treatment. The BBOX radiomics model had high diagnostic performance in the validation cohort, testing cohort (all), and testing cohort (subset).

The BBOX, an easy-to-use tool, can reduce the need for additional labeling and has been used in many clinical investigations (24–28). The region of the BBOX containing both the primary tumor and nearby peritoneum is larger than the ROI of the primary tumor. Recent reports (26, 29–32, 39, 40) have illustrated that peritumoral tissue-based radiomics analysis may reveal valuable information for diagnosis, prognosis and treatment response evaluations. Therefore, in our study, we focused on both the characteristic features of the primary tumor and the peritumoral tissue delineated with the BBOX to develop a radiomics model for the noninvasive diagnosis of OPM.

The analysis of peritumoral tissues surrounding the tumor mass can reveal important information related to tumor aggressiveness; it can reflect lymphovascular invasion, lymphangiogenesis, and angiogenesis (41–43) and provide other information that can be used for diagnostic and prognostic predictions (26, 29–32, 39–42). Moreover, such information may be effectively captured by radiomics analysis (26, 29–32, 39, 40). In our study, the nonoverlapping area between the tumor and R_BBOX was regarded as the PERI, which is theoretically a part of the whole peritumoral region. Our results indicated that the predictive value of PM based on the peritumoral region alone was limited, worse than that of the primary tumor alone, but showed significant improvement when integrated with the ROI of the primary tumor. This may be because the primary tumor, the main disease component, theoretically holds more information about tumor phenotypes than that of PERI, and BBOX covers more regions than the mass or peritoneal region alone.

Few studies (23–28) have investigated radiomic applications for PM diagnosis using CT. Liu et al. (21) and Kim et al. (22) focused on the preoperative CT texture features of the primary tumor area and the omentum area, respectively, to evaluate the possibility of PM in AGCs instead of considering the primary tumor and the surrounding tissue. Dong et al. (23) conducted a radiomics study using the CT phenotypes of primary tumors and the nearby peritoneum to accurately predict OPM in AGC patients. In their study, the diagnostic value of either the peritoneum or the mass was worse than the merged value for the detection of PM, which is consistent with our conclusions. However, they chose regions near the peritoneum rather than the primary tumor, which may be relatively far away from the mass. The delineated ROI of the peritoneum was part of the peritumoral tissue, and the annotated peritoneal area was >2 cm (2), which might not be representative of the entire surrounding condition of the mass on 2D CT images. In addition, manual sketching is a time-consuming and laborious process. The peritoneum is a wide and irregular structure without clear boundaries and forms multiple reflex structures (such as ligaments). Therefore, it is impossible to outline the whole peritoneum. This previous study also selected the local peritoneum, a small area. Although the researchers provided a rationale for selecting this method, delineating the peritoneal area of interest is still very subjective, and thus, it is difficult to determine reproducibility. We used the BBOX method to include all the information on the primary lesion at the 2D level and the information on the peritoneal area surrounding the lesion. These peritoneal areas were not selected subjectively; rather, they were selected incidentally based on their location relative to the primary lesion. Therefore, we believe that our delineation model is simple, easy to use, and reproducible.

Our results showed that by merging the ROI with the PERI, the radiomics model could achieve a higher AUC. This presented a question: what BBOX size would achieve satisfactory diagnostic value for identifying OPM? When the size of the BBOX increased by 10 mm, 15 mm, or no limit, the performance of the models worsened, which might be because the larger the BBOX is, the more tissues are covered. The tissues far away from the carcinoma provided less or no information about the neoplasms. Considering the lower annotation cost and similar AUC of the M_BBOX model compared with the R_BBOX model, the M_BBOX radiomics model was ultimately selected as the final radiomics model.

Laparoscopy remains a useful procedure for evaluating PM status (44–46), but it is invasive and expensive, and its specific applications for GC remain controversial. Our results revealed that the clinical model could not accurately predict PM status. The value of the radiomics model was then assessed by DCA. If the predicted probability of OPM is between 10 and 45%, more cases of undetected OPM on conventional CT images can be detected by the radiomics model than by the all-laparoscopy or no-laparoscopy scheme, avoiding unnecessary surgical procedures and extra costs. Furthermore, patients deemed to have a high possibility of OPM by the radiomics model may undergo diagnostic laparoscopy for further identification, helping to guide proper treatment.

There are several limitations to this study. First, we focused on 2D slices of lesions rather than whole lesions. In future work, automatic segmentation of the whole tumor is worth developing to better predict OPM status. Second, our study was retrospective in nature, and some initially available clinical factors were analyzed. Third, a small number of patients from two external centers were used to evaluate the generalizability of the radiomics model; however, large independent patient cohorts are still needed to validate our results.

The radiomics model based on M_BBOX had higher diagnostic performance for the preoperative detection of OPM in AGC than the radiomics models based on tumor tissue or peritumoral tissue alone. In addition, its annotation is simpler and less time consuming.

The raw data supporting the conclusions of this article will be made available by the authors, without undue reservation.

The studies involving human participants were reviewed and approved by Ethics Committee on Biomedical Research, West China Hospital of Sichuan University. The patients/participants provided their written informed consent to participate in this study.

DL: data curation, software, formal analysis, validation, investigation, methodology, and writing original draft. WZ: data curation, formal analysis, validation, investigation, methodology, and writing original draft. FH: resources, data curation, software, formal analysis, validation, investigation, and writing original draft. PY: data curation, software, formal analysis, validation, investigation, methodology, and visualization. XZ: resources, data curation, software, formal analysis, and validation. HY: conceptualization, data curation, software, formal analysis, validation, and methodology. LY: data curation, formal analysis, and investigation. XF: data curation, formal analysis, and investigation. BS: resources, data curation, formal analysis, supervision, and methodology. BW: resources, data curation, formal analysis, supervision, and methodology. JH: conceptualization, resources, data curation, formal analysis, supervision, funding acquisition, validation, investigation, methodology, project administration, and writing review and editing. ZH: conceptualization, resources, data curation, formal analysis, supervision, validation, investigation, methodology, writing original draft, project administration, and writing review and editing. All authors contributed to the article and approved the submitted version.

This work was supported by the 1*3*5 Project for Disciplines of Excellence, West China Hospital, Sichuan University (No. ZY2017304).

Authors PY and HY were employed by company Infervision.

The remaining 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.