This study developed and validated radiomics models that preoperatively differentiated inverted papilloma from chronic rhinosinusitis with polyps based on unenhanced CT images. It has been shown that RF classifier exhibited higher AUC and accuracy than other classifiers in the validation sets.Recently, Girdler et al. developed a convolution neural network (CNN) model based on nasal endoscopic images to classify IP, CRS, and normal findings which achieved an overall accuracy of 0.74211, but the method was invasive. Woodruff et al. demonstrated that the CT findings of IP were variable and non-specific, when it related to the appearance of benign bony changes, the possibility of potential IP would be considered20. However, not all cases with IP have bony architectural changes. In clinical practice, although contrast-enhanced CT is also helpful in distinguishing these two diseases, there are some risks induced by contrast agents such as non-specific blood pool distribution, potential renal toxicity, and adverse events21, what’s more, using contrast agents may induce increased radiation22. According to MRI findings, IP demonstrated a slightly hyperintense heterogeneous appearance on T1-weighted imaging and intermediate signal intensity on T2-weighted imaging23, while not all cases underwent MRI and the expense was higher than that of CT. Hence, a non-invasive, more convenient, and economical technique to distinguish these two diseases is needed.Our study is the first attempt to differentiate between CRSwNP and IP using radiomics based on unenhanced paranasal sinuses CT images which achieved promising differential performance. Compared with sinonasal endoscopy, contrast-enhanced CT, and MRI, an unenhanced paranasal sinuses CT scan is convenient and facilitative, furthermore, the cost and risk are lower. Therefore, our proposed radiomics models based on unenhanced paranasal sinuses CT images will be a useful tool for differentiating IP from CRSwNP and can be widely applied in clinical tasks.In this multicenter study, four commonly used classifier models (DT, SVM, RF, AdaBoost) were applied to evaluate the performance for discrimination. Impressively, almost all classifiers demonstrated outstanding performance in distinguishing IP from CRSwNP with satisfactory accuracy, among which RF obtained the best results with an AUC of 0.998 (95% CI 0.996–1) in the training set, 0.943 (95% CI 0.903–0.984) in the internal validation set, and 0.934 (95% CI 0.879–0.990) in the external validation set which may be the most optimum model for clinical use. The outstanding performance of RF has also been reported in various other diseases. For example, diagnosis of malaria24, prediction of sepsis after liver transplantation25, and classification of neuroimaging26. Among these diseases, the RF model exhibits optimal performance. As an ensemble machine learning model, RF improves its classification by aggregating multiple (usually hundreds) decision trees using bagging methods. It has the advantages of fast training speed with large samples, small model variance, and strong generalization ability27. Specifically, due to the bagged nature of RF’s stronger resistance to noisy samples and observations, it is more stable than other models in distinguishing IP from CRSwNP.Furthermore, the promising distinguished ability between IP and CRSwNP of our processed radiomics models was not only shown in the training set and internal validation set, but also in the external validation set which reflected the good generalization of this study. However, since IP is mainly unilateral while CRSwNP is mostly bilateral, the promising performance of our radiomics models had the potential of being exaggerated, as we developed the models regardless the lesions were unilateral or bilateral. Importantly, to confirm the accurate differential ability of our proposed differential approach, we further selected the unilateral lesions of both CRSwNP and IP for a subgroup analysis, which also achieved outstanding performance. Additionally, as a comparison, a senior expert’s differential diagnosis results were added with an accuracy of 86.167% by using the data set 1. Consequently, our proposed approach can be generally used as a differential tool for distinguishing IP from CRSwNP based on unenhanced CT images no matter whether the lesion is unilateral or bilateral, which can effectively save time and effort in clinical tasks. Moreover, it can also be used as an aid to expert diagnosis.In total, 43 features from the CT images were extracted, consisting of 9 first-order statistical features, 11 GLRLM features, 20 GLSZM features, 1 neighborhood gray-tone difference matrix feature, and 2 shape features. These findings accentuated the significance of comprehensive features of the lesion region and references to the microscopic and macroscopic features of the lesion. Additionally, the bulk of these features cannot be observed and quantified by human naked eye, which emphasizes the privilege of using automated methodology and extraction of high-throughput radiomics features to aid radiologists and clinicians in disease diagnosis and clinical decision making. However, rather than creating “a black box”28, it is significantly needed for future practice to focus on “explainable AI”29 to explain the implication of radiomics features and the biological behavior behind them17.Presently, texture analysis has been widely applied in the diagnosis and prediction of various diseases, as it can reflect tissue characteristics, among which higher short run low gray level emphasis (SRLGLE) has attracted widespread attention30,31. SRLGLE is a feature used to indicate the distribution of long runs and the joint distribution of short runs and low gray-level values32. Our study found GLRLM-SRLGLE was the best-performing radiomics feature for differentiating IP from CRSwNP, which indicated the difference in histological characteristics of lesions in IP and CRSwNP. Additionally, our subgroup analysis demonstrated that the best feature for the differential diagnosis of IP and CRSwNP was the GLSZM-gray-level non-uniformity (GLNU), which was a measure of the similarity of gray-level values throughout the image33.We attempt to elucidate the pathological characteristics reflected by these two radiomic features. CRSwNP is comprised of highly polypoid mucosa with edema34. The epithelium consists of pseudostratified columnar ciliated epithelium, induced by chronic external stimuli. Beneath the epithelium, there is edematous loose connective tissue infiltrated by inflammatory cells, with eosinophilic infiltration being a distinct histological feature of type 2 endotype. IP demonstrates typical histopathological features of multilayered epithelium invagination into the stroma35. Squamous epithelium is common, but there may also be transitional or columnar epithelium, or a combination of these cell types. The basement membrane is usually intact, and the stroma is richly vascularized with lymphocyte and plasma cell infiltration. Overall, CRSwNP exhibit more severe tissue edema with abundant inflammatory cells infiltration and a coarser texture, whereas IP have richer stromal vascularization with inverted squamous epithelium and a finer tissue texture. This is reflected in the radiomic features, where CRSwNP have lower GLRLM-SRLGLE values (Supplementary Table 3). Song et al. demonstrated that SRLGLE is an independent predictor for identifying lymph node infiltration status in colorectal cancer, with rougher lymph node texture and a significantly decreased SRLGLE value in N1-2 patients32. In contrast, the cellular arrangement and types differ, GLSZM-GLNU likely reflects the degree of inflammation and edema in the tissues. Our subgroup analysis revealed higher GLSZM-GLNU values in highly edematous and inflamed CRSwNP (Supplementary Table 4). This is consistent with the findings of Baessler et al. who observed higher GLNU values in acute heart failure myocarditis compared to chronic heart failure myocarditis, indicating a greater degree of myocardial edema and inflammatory response36. Parvaze et al. also found higher GLNU values in high-grade gliomas compared to brain metastases, reflecting a more severe degree of brain edema37. Collectively, SRLGLE and GLNU were identified as the strongest differential factors in our analyses respectively, the use of these imaging biomarkers could help clinicians make accurate disease diagnoses.However, larger amounts of data and multicenter tests are needed for further demonstrating and optimizing our differential model before implementing it in clinical practice to establish an AI differential diagnosis platform, which will fast and accurately distinguish these two diseases when clinicians or patients import digital CT into this platform.Despite promising differential models, our study still has several limitations. First, the paranasal sinuses CT images were from a variety of different CT scanners, although the images normalizing before extracting the features was done, our study still has the potential for bias in the analysis. Next, all ROIs were manually segmented in each slice of the sequences of CT, which was time-consuming. We have proposed a method of automatic sinuses segmentation incorporating handcrafted segmentation in another study, which may be applied to the differentiation of IP from CRSwNP and decrease the workload of sinuses segmentation in the future38. Finally, more data from more centers should be collected to further confirm the accuracy of our proposed method and to establish an AI differential diagnosis platform.In conclusion, this multicenter study showed the outstanding differential ability of CT-based radiomics in distinguishing inverted papilloma from chronic rhinosinusitis with polyps, which could help clinicians make accurate, non-invasive, and economical disease diagnoses and promote the development of precision therapy.
