Accurate differentiation between sinonasal IP and NP has always been a significant challenge in clinical practice. The resemblance in clinical symptoms and the results of various auxiliary examinations contribute to the difficulty in distinguishing between these conditions [16]. Given that IP typically requires more extensive surgical resection, whereas NP can be managed with medical treatment or minimally invasive surgery, accurate preoperative prediction of IP or NP is crucial for determining the optimal surgical strategy and predicting the prognosis.

In the field of otolaryngology, clinicians commonly employ CT scans prior to surgery to assess the extent of IP involvement in the sinuses and determine the appropriate surgical strategy. Earlier studies have indicated that focal hyperostosis is a notable CT feature for predicting IP [17]. For instance, Glikson et al. conducted a study involving 195 IP patients and discovered that 65% of them exhibited focal hyperostosis [18]. However, the study concluded that the presence of focal hyperostosis detected in preoperative CT scans did not significantly impact the long-term prognosis of inverted papilloma resection. These findings collectively suggest that relying solely on the presence of focal hyperostosis is insufficient for distinguishing between IP and NP. Another study by Sukenik et al. reported a specificity of only 20% for preoperative CT examination in evaluating IP [19]. Similarly, in our own study, a considerable number of IP patients did not exhibit focal hyperostosis. These research findings highlight the insufficiency of focal hyperostosis alone in differentiating IP from NP.

Nevertheless, the potential of CT in distinguishing the two remains to be tapped. Sano et al. tried to evaluate the meaning of relative CT value (CT attenuation number relative to those of the brainstem) for distinguishing NP and IP [15]. According to this research, CT values of the NP set and IP set were 28.0 ± 12.8 HU and 38.0 ± 10.0 HU separately, suggesting IP had superior CT values than NP. Therefore, we could acquire a simple and creative reference parameter for diagnosis of NP and IP. Based on this, our study tries to find a new way to extract image features of two diseases, including CT values, to maximize the use of CT images to achieve the identification of IP and NP, which involves the field of radiomics.

Radiomics currently plays a crucial role in non-invasive clinical diagnosis, particularly in the field of tumors, and has shown promise in classification, staging, predicting healing outcomes of various tumors, and evaluating treatment effects of different surgical options [20,21,22,23,24,25]. Moreover, radiomics can extract informative features from images that may be difficult for humans to detect [26, 27]. As the clinical application of radiomics continues to be explored, an increasing number of otorhinolaryngology practitioners have started applying radiomics to their field. Yan et al. demonstrated that a model combining morphological features and MR radiomics could accurately predict inverted papilloma with squamous cell carcinoma transformation, potentially enhancing patient consultation and facilitating more precise treatment planning [28]. Studies have also shown the feasibility of using artificial intelligence and radiomics to diagnose difficult cases involving small round cell malignant tumors (SRCMTs) or non-SRCMTs [29, 30]. These findings suggest the promising potential of non-invasive methods in clinical practice. Recent research by Du et al. revealed that a novel combination of clinical features and MRI-based radiomics could effectively differentiate between IP and NP invading the olfactory nerve, highlighting the potential of radiomic models in addressing nasal diseases [31].

Radiomics has demonstrated strong capabilities in feature extraction and texture analysis, making it suitable for the intrinsic appearance evaluation of tumors. Previous studies have shown that radiomics can provide valuable insights, particularly in cases where radiologists have limited recognition ability, enabling precise and automatic tumor extraction [32,33,34,35]. While MRI is known for its ability to display soft tissues, making it potentially more suitable for differentiating between tumors and inflammatory conditions such as IP and NP, in clinical practice, patients suspected of having IP or NP often undergo CT examinations initially due to the high cost of MRI and the limited visualization of bone structures. Therefore, this study focusing on CT-based radiomics is more practical and applicable in a clinical setting compared to MRI-based radiomic research.

In our study, we developed several radiomic models based on CT features to distinguish between IP and NP. The results demonstrated that these radiomic models, particularly the XGBoost classifier, exhibited excellent diagnostic accuracy, confirming their clinical application value. Our study possesses several advantages compared to previous studies on radiomics models for differentiating NP and IP. Firstly, we approached the analysis of IP and NP from a clinical perspective rather than relying solely on a database, and our study included a larger population compared to most relevant studies. Secondly, we employed three different methods of feature extraction (correlation coefficient, Boruta, and random forest), which contributed to improved diagnostic performance compared to previous radiomics models. Thirdly, we utilized multiple classifier models for classification, and some of these models demonstrated outstanding performance. Additionally, our study employed ten-fold cross-validation to validate the accuracy of the algorithm, a step that was lacking in previous studies.

Although our study has achieved some results, there are certain limitations that should be discussed. Firstly, we did not incorporate clinical factors or clinical indicator features into the radiomics analysis. Clinical diagnosis is typically based on the comprehensive evaluation of all available data in clinical practice, including imaging information of surrounding tissues and the impact of the tumor on neighboring structures. Therefore, it is important to conduct further research on clinical indicators and explore their performance and impact on the identification of IP and NP. Fusion experiments incorporating clinical indicators could enhance the diagnostic efficiency of IP.

Secondly, the number of clinical cases in our study was limited, particularly the data on IP patients, which are relatively scarce in hospitals. This limited sample size affects the generalizability of our model. It would be beneficial to collect a larger and more diverse dataset to improve the robustness and applicability of the radiomics models.

Lastly, the machine learning methods employed in our study mainly focused on traditional algorithms and did not involve deep learning techniques. Deep learning has demonstrated significant advancements and achievements in various research fields. Therefore, further investigations utilizing deep learning methods are expected to contribute to the differential diagnosis between IP and NP.