Cervical cancer is the fourth most common female cancer worldwide (1). In 2016, there were approximately 34,000 cervical cancer-related deaths in Chinese women (2). Squamous cell carcinoma (SCC) and adenocarcinoma (AC) are the main pathological subtypes of cervical cancer, accounting for 70-75% and 10-25% respectively. The incidence of AC has been observed to increase in recent decades (3). Patients with locally advanced cervical cancer who receive radiation therapy or concurrent chemoradiotherapy have a worse prognosis for AC compared to SCC, highlighting the need for alternative treatment options specifically for AC cases (4). High intratumor heterogeneity exhibited a significantly poor clinical outcome (5). Therefore, it is essential to uncover the differences between AC and SCC from multiple perspectives, explore the underlying reasons for these differences, and develop personalized treatment strategies and plans, as this holds considerable importance.

Pathological diagnosis is considered the gold standard for the detection of cervical cancer, with cervical cytology and cervical biopsy being the primary recommended methods (6). However, AC may sometimes result in cytological false-negatives (7). Biopsy is an invasive procedure associated with risks of bleeding and infection. Point-to-point biopsy performed on larger tumors only evaluates a small portion of the sample, resulting in sampling bias and an inability to comprehensively assess tumor heterogeneity (8–10). In addition, radiologists find it challenging to differentiate AC from SCC based on conventional imaging modalities such as magnetic resonance imaging (MRI), positron emission tomography (PET)/computed tomography (CT), ultrasound, etc., and the interobserver agreement is typically low (11, 12).

Radiomics is a rapidly growing field of research that utilizes medical images to extract quantitative features, converting them into high-dimensional data for analysis and exploration. This technique enhances our understanding of diseases and provides valuable support for clinical decision-making (13, 14). Malignant tumors exhibit considerable spatial variation within the tumor at the morphological and histopathological levels, including cellularity, vascularization, extracellular matrix, and necrotic components (15, 16). As a noninvasive tool, radiomics can quantify intratumoral heterogeneity and is widely used in diagnosis, treatment response evaluation, and survival prediction (17). Among conventional imaging modalities, MRI has better fine exquisite soft tissue resolution than CT and PET, so it has long been considered as the preferred imaging method of choice for the evaluation of local tumor extension in primary cervical cancer. Unsurprisingly, numerous radiomics studies aimed at identifying the pathological subtypes of cervical cancer are primarily founded on pretreatment MRI (18). To the best of our knowledge, there are no radiomics studies that delineate the primary cervical tumor on CT images, which might be related to the inability to clearly define the boundary of the primary cervical tumor on CT images. In addition, only two published studies have preliminarily evaluated PET radiomics in cervical cancer to discriminate between AC and SCC (19, 20). Nevertheless, both of the studies were single-center ones, extracted too few radiomic features, and the methods for radiomics feature selection and model construction were simplistic. Even in one of the studies, only 83 patients were included.

Previous studies have shown that radiomic features based on CT or PET images can achieve the differentiation of pathological types of lung cancer (21). The application of radiomics methods for diagnosis and tumor characterization might be a potential supplement for omics datasets, or an alternative for pathological diagnosis, particularly for patients who are at an advanced stage, inoperable, or unable to undergo biopsies. A more extensive and comprehensive study is required to investigate the value of PET/CT imaging in differentiating the subtypes of cervical cancer. Therefore, the aim of this study was to develop and validate an optimal machine learning (ML) model based on pretherapeutic fluorine-18-fluorodeoxyglucose (¹⁸F-FDG) PET/CT for differentiating between SCC and AC in cervical cancer.

In this study, we successfully developed six ML models based on PET and CT images, among which the lightGBM model based on PET radiomics features performed excellently in distinguishing AC and SCC.

Previous literature has indicated that CT radiomics features exhibit better predictive performance than PET radiomics features in predicting survival, and CT radiomics features are also more abundant than PET features (22, 23). Nevertheless, with regard to distinguishing pathological subtypes, Kirienko et al. (24) discovered that PET radiomics features had a greater ability to discriminate between primary and metastatic pulmonary lesions than CT radiomics features. Further studies conducted by Hyun et al. (25) and Han et al. (21) demonstrated that a PET/CT-based machine learning method was able to make a distinction between AC and SCC in patients with non-small cell lung cancer. This study indicates that in distinguishing SCC and AC, the selected PET radiomics features are substantially more numerous than CT radiomics features, and the performance of the PET radiomics model is notably superior to that of the CT radiomics model. Furthermore, the Delong test showed that although there was a slight improvement in performance when combining PET radiomics features with CT radiomics features, the increase in AUC value did not reach statistical significance (the p values of the training and internal validation cohorts were 0.347 and 0.776, respectively). In a retrospective study, Shen et al. (19) first found that short-zone emphasis (SZE) from GLSZM was the only PET-based radiomics feature that showed quantitative differences between SCC and non-SCC in cervical cancer. Tsujikawa et al. (20) reported that the correlation from normalized gray-level co-occurrence matrix (NGLCM) was the only feature extracted from ¹⁸F-FDG PET that showed significant differences between cervical SCC and non-SCC. The previous two published studies extracted merely 18 or 76 features from the original images. In contrast, our study extracted 2818 features from the original images as well as the converted images. This might be the cause for which the previous study could select only one meaningful feature, while our study selected five. In summary, the findings of this series of PET/CT radiomic studies highlights the importance of functional imaging-based radiomics research in differentiating tumor pathological subtypes. This may be related to the FDG uptake heterogeneity between AC and SCC, which is consistent with the identification of the pathological subtypes of lung cancer based on PET/CT (21).

MRI techniques also introduced various functional sequences, including apparent diffusion coefficient (ADC), dynamic contrast-enhanced imaging, and perfusion-weighted imaging (26). Wang et al. (18) achieved good differentiation between SCC and AC using a multiparameter MRI radiomics model based on ADC, enhanced T1-weighted imaging, and other anatomical and functional sequences. Although the differentiation performance of the multi-parametric MRI-based radiomics model was the highest among the published MRI-based radiomics studies, its differentiation performance (AUC = 0.89) was lower than that of the pure PET-based radiomics model constructed in our study (AUC = 0.955). These findings demonstrate the advantages of PET radiomics features over multiparametric MRI radiomics features to a certain extent. PET-based radiomics can not only reveal the intratumoral heterogeneity of imaging structures between AC and SCC but also demonstrate the heterogeneity of tumor cell metabolism. Moreover, radiomics features are based on manually segmented ROIs in five MRI sequences, which not only requires a substantial amount of work but also increases the instability of the features.

The radiomics features selected in our study are all derived from processed images, which may reveal greater tumor heterogeneity differences between SCC and AC compared to the original images, showcasing the advantage of radiomics. Among these radiomics features selected in our study, the firstorder_Maximum and firstorder_Minimum represent the maximum and minimum gray level intensities, respectively. SCC exhibits significantly higher values than AC, indicating that SCC has a stronger FDG uptake than AC. Campos-Parra et al. found that compared to AC, SCC exhibits higher activation levels of key cancer pathways, such as IL-17, JAK/STAT, and Ras signaling (27). high-risk human papilloma virus (HPV) -16 infection is more common in SCC, while HPV-18 and HPV-45 are more frequently observed in AC (27, 28). Priego-Hernández et al. discovered that cervical cancer and HPV-16-positive cell lines have increased expression of HIF-1α and glucose metabolism-related genes (GLUT1, LDHA, CAIX, MCT4, and BSG) (29). Furthermore, there are significant variations in the expression of glucose metabolism-related genes between SCC and AC (30). Choi et al. demonstrated that tumor FDG uptake is associated with glucose transporters (Glut-1 and Glut-3), with SCC exhibiting higher expression intensity and proportion of Glut-1 compared to AC. Consequently, SCC demonstrates higher SUVmax and stronger FDG uptake capacity (31). Small dependence emphasis (SDE) from GLDM and small area low gray level emphasis (SALGLE) from GLSZM represent tumor heterogeneity, with higher values indicating more significant heterogeneity. In our study, GLDM_SDE and GLSZM_SALGLE features were significantly higher in SCC compared to AC, indicating that the intratumoral metabolic heterogeneity based on PET imaging in SCC is significantly higher than that in AC. This may be related to the previously mentioned metabolic and histomorphological differences between SCC and AC. The tissue structure of SCC is tight, with small gaps between tumor cells, wrapped in several matrix structures, forming cancer nests. In contrast, the tissue structure of AC is more loose, characterized by glandular differentiation. Therefore, the differential expression of pathogenic molecular mechanisms, especially glucose metabolism genes, determines the metabolic differences of tumor cells, while cell arrangement and tissue morphology determine the spatial heterogeneity of tumor cells. The tumor heterogeneity revealed by PET images manifests these metabolic differences and spatial heterogeneity of tumor cells. These findings require further validation with a larger-scale patient or in combination with pathomics.

In this study, we employed six ML algorithms to develop models for distinguishing SCC and AC. Among the algorithms, the radiomics model constructed by the LightGBM algorithm exhibited excellent differentiation performance, accuracy, sensitivity, and specificity with a relatively balanced performance. This finding is consistent with a similar study conducted by Lam et al., who investigated the correlation between radiomics features and tumor mutation burden in glioma based on MRI images using LR, SVM, and six other ML algorithms (32). They found that the radiomics model constructed by the LightGBM algorithm also demonstrated the best discriminative performance with relatively balanced sensitivity and specificity. Furthermore, researchers have successfully achieved good discriminative performance in distinguishing low-grade and high-grade meningiomas using the LightGBM algorithm for both radiomics and deep learning models (33). Similarly, Chang et al. constructed LightGBM and convolutional neural network (CNN) models based on noncontrast CT and enhanced images to differentiate thymic epithelial tumors from other anterior mediastinal tumors (34). The results demonstrated that the LightGBM model outperformed the CNN model in both the noncontrast CT dataset and the enhanced CT dataset. The LightGBM algorithm, which is based on the gradient boosting decision tree model, optimizes the search for optimal split points and the tree growth process. It supports efficient parallel training and possesses advantages such as faster training speed, lower memory consumption, better accuracy, and quick processing of massive data, making it widely applicable. Therefore, ML can better handle complex nonlinear relationships in large-scale datasets and holds great potential for clinical applications (35). However, it is important to acknowledge that ML models and algorithms also have limitations, including overfitting and lack of interpretability. Overfitting can undermine predictive performance, while the lack of interpretability can hinder the use of ML (36). Hence, it is essential to prioritize the future optimization of ML algorithms and conduct independent validations to verify their performance.

There were several limitations in this study. Firstly, it was a retrospective and preliminary study, carrying a potential selection bias despite the use of strict inclusion and exclusion criteria. Secondly, HPV status and histological differentiation were not available for some patients when retrieving the electronic medical record system, and we were unable to further explore their impact on pathological subtypes. Lastly, the sample size of AC in this study is relatively small, but this is consistent with the epidemiology of cervical cancer. To improve the generalizability of the model, it is necessary to investigate a larger sample size from multiple centers in future research.

The lightGBM-based PET radiomics model effectively identified pathological subtypes in patients with locally advanced cervical cancer and may help clinicians in their daily decision-making process.