Pituitary adenoma is one of the most prevalent intracranial masses that can affect adults (1, 2). The varied clinical manifestations usually result from the endocrine activity, or volume of tumors. The classification of PA is based on different criteria, such as size, immunohistochemistry (IHC), invasion, hormone secretion, and clinical manifestation (1, 3). PAs are classified into micro, macro, and giant adenomas by the MRI size. The IHC subtypes of PAs are composed of growth hormone (GH), prolactin (PRL), adrenocorticotropic hormone (ACTH), thyroid-stimulating hormone (TSH), and follicle-stimulating hormone-luteinising hormone (FSH-LH), including the monohormonal and plurihormonal adenomas. Although benign in terms of their biological behavior, 30–45% of tumors invade the cavernous or sphenoid sinus, which can be categorized into invasive and non-invasive adenomas (1, 3, 4). According to the clinical classification, PAs consist of functioning and non-functioning types (1).

The treatment strategy for most tumors is operation. The postsurgical recurrence rate of pituitary macroadenoma (PMA) within 5 years is considerably high (5). The tendency to relapse has been related to many factors, including different histotypes, tumor remnants, or the extent of invasion into adjacent anatomical structures (6). Previous research has demonstrated that many clinicopathological prognostic tools have potential to predict the recurrence of PMA, consisting of IHC characteristics, invasion of tumors, genetic expression, and markers of proliferation (7–11). However, very few attempts have been made to integrate these risk factors with a machine-learning approach.

Radiomics is a form of analysis that quantitatively extracts imaging features from medical data (12). Thus far, radiomics studies of PA have predominantly focused on two aspects: presurgical evaluation and subtype classification (13–17). Some researchers have used radiomics to explore the potential for relapse in PA. However, these previous models have been associated with small sample sizes and only cases involving non-functioning tumors (18, 19).

In the present study, we aimed to establish a comprehensive classification model that combined independent clinicopathological risk factors with preoperative radiomics signatures for the prediction of PMA recurrence within 5 years of surgery. Our goal was to provide an efficient tool for guiding clinical management and predicting prognosis.

In this study, we constructed a diagnostic classification strategy based on the presurgical MRI to predict the recurrence risk of PMA within 5 years. This comprehensive classifier incorporates clinicopathological and radiomics features and can accurately predict the recurrence of PMA.

Various factors are known to be associated with a higher risk of PMA recurrence, which remains a significant problem for both clinicians and patients. Very few previous reports have attempted to combine clinicopathology analysis with radiomics for the prediction of PMA. MRI radiomics approaches have been described in previous literature. For example, Zhang et al. (18) and Machado et al. (19) used this method to explore the recurrence of non-functioning PA. Compared with these previous results, the Model 2 in our study presented with relative low diagnostic accuracy. This is probably because we built the test set using an independent set of subjects rather than the training group. Moreover, our study simultaneously included functioning and non-functioning PMAs. The study cohorts and enrollment criteria may lead to the different predictive performances. However, our Model 2, which had a relatively large sample size and incorporated comprehensive markers, showed a better level of classification performance than Model 1. This improved predictive efficiency demonstrates that the combination of clinicopathological data and imaging may provide more practical information and guidance for developing a treatment and prognosis strategy than clinical analysis alone.

MLP model is a feed-forward artificial neural network (ANN) model that is applicable to a non-linear inseparable issue; Almubark et al. demonstrated the predictive value of this approach in their previous study (26). The generalization and efficacy of this method have been widely confirmed in several papers (27–29). Given these characteristics, we also established an MLP classifier for the recurrence of PMA and achieved satisfactory levels of predictive performance in a test cohort. These data indicate that this deep learning algorithm is a reproducible and robust technique for classification.

Many risk factors are associated with the recurrence of PMA. In the present study, we incorporated some of the primary predictors that have been described in previous literature (30). Four clinicopathological risk factors—age, height, invasion, and residual tumor—were finally included in our comprehensive model; these factors were identified by a combination of univariate and multivariate analyses. We found that the patients in the recurrence group tended to be younger. We believe that this is because there is a greater risk of gene disorders in younger age groups (31). In a previous study, Trott concluded that young patients express elevated levels of ki-67 in non-functioning pituitary adenoma, and that this is strongly associated with relapse (32). Moreover, some ultrastructural types of PMA resulted in regrowth, such as sparsely granulated somatotroph adenoma, and are more likely to affect younger patients. The aggressive growth pattern of PMA is one of the main reasons concerned with the prognosis (33). The invasive tumors usually exhibit more rapid growth, a higher proliferative index, and larger size. Thus, the severe erosion of surrounding structures (e.g., cavernous sinus and sellar floor) and great extension of the supra- and para-sellar lead to increased rates of recurrence (34). Similar results in the present literature provide support to our conclusion that tumor remnant is also known to be significantly correlated with PMA behavior, especially the higher incidence of larger extra-sellar residuals (5, 6, 34). Height is another recognized predictive candidate. It is evident that the tumors with higher height may result in incomplete resection and invasive behavior, which consequently raise the likelihood of regrowth. These findings are consistent with previous reports (5, 6, 35). This suggests that the clinical characteristics described above are useful and reliable tools for predicting the prognosis of patients with PMA.

The tumor classification involving the transcription factor or ultrastructure showed that the presence of giant lactotroph, sparsely granulated somatotroph, crooke's cell, or silent corticotroph adenomas tends to present the recurrence nature (5, 36, 37). Our study focused on the proposed IHC subtypes that are also important and potential indicators related to the progression. The study reported by Asioli et al. showed that PRL, ACHT, and FSH-LH subunits had relapse risk with high probability (20). These indices were not statistically significant when compared between the two groups in our study; it is possible that this was owing to the small sample size compared with the previous study. Although the rate of TSH adenoma was the lowest among all cases, the trend of incidence is in line with tangible clinical practice. The morbidity of this type is low based on the demographic investigation, comprising <3% of all tumors (38). We included the plurihormonal adenoma in our study. The most common type is the co-secretion of GH and PRL (39). But the combination of different hormones tends to be more complicated. Little is known with regard to the correlation between recurrence and plurihormonal tumors.

The trans-sphenoidal and endoscopic surgical methods are extensively applied to dealing with PMA (40–42). Our study did not show the correlation between operative approaches and recurrence. According to current studies, tumor size and invasive extension were decisive factors in extent of resection. The macro or giant adenomas and extensive invasion tended to be difficult to achieve grossly complete resection, although the trans-sphenoidal or endoscopic resection was used (43–45). The residual tumor is likely to be the crucial factor that affects the prognosis, suggesting that the surgical resection may be more likely to associate with intrinsic biological characteristics of tumors, compared with an operative procedure. Patients who received total resection still have the possibility of relapse (5). This indicates that surgery alone may not enable to decrease the recurrence rate. It is also important and beneficial to combine radiotherapy or other methods (46).

The consistency of adenomas probably influences the prognostic outcomes. A study showed that texture was correlated with tumor profiles, complications, and surgical resection. The hard adenomas were at higher risk of large and aggressive behavior and subtotal removal (47). This suggests that consistency is potential to predict the recurrence of PMAs. Rui et al. confirmed the utility of the radiomics method for determining the texture of PMAs (13). Future PA studies may pay more attention to the relationship between the stiffness and recurrence by radiomics.

The proliferative biomarkers of ki-67, p53, and mitosis play an important role in tumor prognosis. Although there remains the controversy of arguments, the changes of these proliferation indices are often associated with aggressive PAs. The prognoses of the patients tended to be recurrent or poorer in the presence of ki-67 ≥3, mitoses >2, and p53 overexpression based on the study by Raverot and European Society of Endocrinology (48). Above all, the most effective and useful predictive strategies are incorporating the predictors of different fields, such as clinical, imaging, and immunohistochemical examination.

First, the proliferations, transcription factors, ultrastructural subtypes, along with expression profiles of certain genes, were not considered in this study but may improve the performance of our classifier. Second, the pituitary scanning sequence with smaller slice thickness and interval was not applied in the study, because the protocol is not clinical routine examination; it may be considered in future studies. Third, microadenoma is another common subgroup of pituitary tumor; radiomic studies of this form of tumor are very rare. We did not include this type of tumor in the present study due to a limited sample size. Finally, this study was based in a single center and lacks external validation in multiple centers.

The combination of clinicopathological characteristics and imaging is useful for predicting the recurrence of PMA within 5 years. The integrated classifier was superior to a clinical classifier and may facilitate the prediction of individualized prognosis and therapy.

The original contributions presented in the study are included in the article/Supplementary Material, further inquiries can be directed to the corresponding author/s.

The studies involving human participants were reviewed and approved by the Ethics Committee of Beijing Tiantan Hospital. The Ethics Committee obtained written informed consent from the participants.

JM designed the original research. YZ conducted the research and wrote the manuscript. YZ, YLu, and XK analyzed the data. JM, TW, and YLo revised the paper. All authors contributed to the article and approved the submitted version.

This research project was funded by National Natural Science Foundation of China (No. 61771325).

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.