Advancing fungal sinusitis diagnosis: a radiomics and machine learning approach

Abstract

Background

Differentiating fungal from chronic sinusitis remains a diagnostic challenge due to overlapping symptoms.

Objectives

This study aims to enhance diagnostic accuracy for fungal sinusitis by integrating radiomic analysis with machine learning.

Methods

Data from 106 fungal and 146 chronic sinusitis patients, confirmed by surgical pathology at the Integrated Traditional Chinese and Western Medicine Hospital of Wenzhou Medical University from January 2022 to December 2025, were analyzed. Radiomic features from CT scans were extracted and reduced to 19 key indicators using 3DSlicer, Minimum Redundancy Maximum Relevance (mRMR), and lasso regression. A comparative analysis of logistic regression, support vector machine, and random forest models selected the optimal model based on AUC performance.

Results

Using mRMR followed by LASSO regression, 19 key radiomic features were selected to distinguish fungal sinusitis from chronic sinusitis. Among logistic regression, SVM, and random forest, the random forest model performed best, achieving a training AUC of 96.04% and a test AUC of 94.91%, with corresponding accuracies of 92.51% and 91.30%. Calibration curves confirmed excellent agreement between predicted and actual outcomes, demonstrating strong diagnostic reliability. Calibration curves showed strong agreement between predicted probabilities and actual outcomes.

Conclusions

Employing radiomic features and machine learning, specifically a refined random forest model, significantly improves fungal sinusitis diagnosis accuracy. This approach promises to minimize diagnostic errors and enhance therapeutic decisions, marking a significant advance in precise diagnostics for fungal sinusitis. However, external validation from independent institutions was not performed, representing a key limitation of the study.

Clinical trial number

Not applicable.

Supplementary Information

The online version contains supplementary material available at 10.1186/s12880-026-02494-x.

Introduction

Fungal Sinusitis, an inflammatory condition of the sinuses induced by fungal infections, has seen a rising incidence over recent years. Previously considered rare, its increasing prevalence is thought to be closely linked with the widespread use of antibiotics, heightened utilization of immunosuppressive medications, and an enhanced awareness and understanding of the disease [1–3]. The pathogenesis of fungal sinusitis is complex, with clinical manifestations that vary widely, leading to frequent misdiagnoses as other forms of sinusitis. This misidentification poses significant challenges for accurate diagnosis and timely intervention, underscoring the need for more precise diagnostic tools [4, 5].

Despite its relatively lower incidence in comparison to bacterial sinusitis, the potential for severe complications, including ocular and intracranial complications, necessitates timely and accurate diagnosis of fungal sinusitis [6]. Traditional diagnostic methodologies rely heavily on clinical manifestations, radiological findings, and histopathological confirmation. However, the overlap in clinical and radiological features between fungal sinusitis and more common forms of sinusitis often complicates diagnosis. While histopathological examination remains the gold standard for diagnosis, its invasive nature, coupled with time constraints, renders it an impractical option for all patients [7].

The integration of radiomics with machine learning offers a promising solution to this diagnostic dilemma. The fundamental premise is that fungal sinusitis may exhibit subtle, visually imperceptible patterns on CT images, such as specific textural heterogeneity or density variations caused by fungal elements [8]. Radiomics can quantify these patterns by extracting a high-dimensional set of minable features from standard medical images. Subsequent analysis using ML algorithms can discern complex patterns within this data to build a robust diagnostic classifier [9]. While the potential of this synergy is recognized, its specific application for differentiating fungal sinusitis from chronic sinusitis has not been systematically investigated. Most radiomics studies have focused on broader sinonasal disease categories, whereas fungal and chronic sinusitis share overlapping imaging features and may coexist. Their differentiation is clinically relevant because treatment strategies differ: surgery plus antifungals for fungal sinusitis versus medical management for chronic sinusitis. This gap supports the development of a dedicated model for this specific challenge.

To address this gap, our study aims to develop and validate a non-invasive diagnostic model by systematically combining CT-based radiomics with machine learning. We hypothesize that a classifier built on discriminative radiomic features can achieve superior diagnostic accuracy compared to conventional radiological assessment. Utilizing data from 252 pathologically confirmed cases, we will extract radiomic features from sinus CT scans using a standardized pipeline involving 3D-Slicer for segmentation and the pyradiomics library for feature extraction. A rigorous feature selection process, combining mRMR and Lasso regression, will be employed to identify the most predictive features and mitigate overfitting. We will then implement and compare three classical ML models—Logistic Regression, Support Vector Machine (SVM), and Random Forest—to determine the optimal algorithm for this task, with model performance rigorously evaluated by metrics including AUC, accuracy, and recall. We anticipate that this methodology will yield a reliable tool that can assist clinicians in making accurate, early diagnoses, thereby improving patient management strategies for fungal sinusitis.

Materials and methods

Study design and patient selection

This investigation was a retrospective study carried out at the Integrated Traditional Chinese and Western Medicine Hospital of Wenzhou Medical University, encompassing patient data from January 2022 to December 2025. Institutional Review Board approval was secured prior to commencement, with a waiver for informed consent granted due to the retrospective analysis nature. The study population consisted of 252 individuals, including 106 patients diagnosed with fungal sinusitis confirmed via surgical pathology, and 146 control patients with chronic sinusitis. The dataset was first split into training (80%) and test (20%) sets, after which all preprocessing steps (including normalization) were performed using parameters derived exclusively from the training set; feature selection using mRMR and LASSO with 10-fold cross-validation was then applied strictly within the training set, followed by hyperparameter tuning using grid search with 10-fold cross-validation on the training set only; finally, the fully trained model was evaluated once on the untouched test set.

Inclusion criteria mandated: (1) receipt of a pre-surgical sinus CT scan within one month of the operation, (2) CT evidence of increased density within the sinuses, and (3) availability of complete clinical and diagnostic data.

Exclusions were applied for (1) substandard CT imaging unsuitable for detailed radiomic analysis, (2) history of sinus surgery within six months prior to the study period, and (3) insufficient medical documentation.

Clinical and conventional imaging data collection

To characterize the study cohort and enable comparison with existing literature, we collected baseline clinical and conventional imaging data for all enrolled patients. This included demographic information (age, sex), clinical symptom duration, laterality of disease (unilateral or bilateral), and the specific sinus groups involved (maxillary, ethmoid, sphenoid, frontal). Furthermore, key conventional CT imaging features suggestive of fungal sinusitis were documented, including the presence of intrasinus hyperdensity, bony remodeling/sclerosis, and bony erosion. Categorical variables were compared using the Chi-square test, and continuous variables were compared using the independent Student’s t-test or Mann-Whitney U test, as appropriate.

CT imaging and radiomic feature extraction

CT imaging was standardized across the cohort using a 64-slice multidetector CT scanner (SOMATOM Sensation 64; Siemens, Erlangen, Germany). Parameters were set for a 0.5 mm slice thickness and a 200 mm field of view, with all images exported in DICOM format for subsequent analysis. The 3D Slicer software (version 4.11.20210226, www.slicer.org) was utilized for image reconstruction and the precise annotation of regions of interest (ROIs) within the sinus cavities. For all patients, ROIs were manually delineated slice-by-slice on axial CT images using a soft-tissue window (window width/level: 400/40 Hounsfield Units). The ROI encompassed the entire abnormal content within the affected sinus cavity. Specifically, for the maxillary sinus, the ROI included all areas of soft-tissue attenuation or fungal ball within the sinus boundaries, while carefully excluding the bony sinus walls and air spaces outside the sinus ostium. In cases of diffuse mucosal thickening, the boundary was defined as the interface between the abnormal mucosa and the air lumen or adjacent bony wall. To ensure annotation accuracy and mitigate inter-observer variability, two radiologists with extensive experience independently delineated the ROIs. 30 randomly selected cases (20% of the cohort) were independently segmented by both radiologists, and the intraclass correlation coefficient (ICC) was calculated for all extracted features. Features with ICC < 0.75 were considered unstable and excluded from subsequent analysis, while only robust features (ICC ≥ 0.75) were retained for feature selection (mRMR and LASSO) and model building. Of the 129 initially extracted radiomic features, 112 demonstrated excellent inter-observer agreement (ICC ≥ 0.75, range: 0.76 to 0.98, median ICC = 0.89) and were retained for subsequent analysis. The remaining 17 features with ICC < 0.75 were excluded due to poor reproducibility. For the final model building, consensus segmentation determined by two experienced radiologists was used for all patients.

Radiomic feature extraction and preprocessing

Radiomic feature extraction was performed using the open-source PyRadiomics library (version 3.0.1) in Python, utilizing its default parameter file (Params.yaml) without modification except for the gray-level discretization setting. All features were calculated in 3D using default pixel spacing correction and feature normalization settings. Feature extraction was conducted on the original, non-filtered image using the 3D VOIs. Prior to feature extraction, image preprocessing was applied: First, a global Z-score normalization was applied to the voxel intensities of the entire CT image, using the mean and standard deviation parameters calculated exclusively from the training set. Subsequently, gray-level discretization was performed using a fixed bin width of 25 Hounsfield Units (HU) to standardize intensity scales across different scans.A standardized set of feature categories was extracted, which included: Shape-based features (e.g., Volume, Surface Area, Sphericity), First-order statistics (e.g., Mean, Median, Skewness, Kurtosis of intensity values), Texture features from Gray Level Co-occurrence Matrix (GLCM), Gray Level Run Length Matrix (GLRLM), Gray Level Size Zone Matrix (GLSZM), and Neighboring Gray Tone Difference Matrix (NGTDM), and higher-order features from wavelet transforms (applying wavelet filters to the original image prior to feature calculation).

A total of 129 radiomic features were initially extracted from each patient’s ROI. Prior to feature extraction, image preprocessing was applied. A fixed bin width of 25 Hounsfield Units (HU) was used for gray-level discretization to standardize intensity scales across different scans. Additionally, image normalization (Z-score normalization) was performed, but critically, the mean and standard deviation parameters for normalization were calculated exclusively from the training set and then applied to both the training and test sets.

Feature selection and machine learning model development

The entire modeling workflow was designed and executed with a strict separation between data used for model development and data used for final evaluation to ensure an unbiased performance estimate. The complete dataset (n = 252) was first randomly partitioned into a temporary training set (80%, n = 201) and a hold-out test set (20%, n = 51). The test set was sequestered and remained completely untouched during all subsequent development steps.

All preprocessing, feature selection, and model optimization were conducted exclusively within the temporary training set. Image normalization (Z-score) was applied using the mean and standard deviation parameters calculated solely from the training set data. To ensure feature robustness, features with an intraclass correlation coefficient (ICC) below 0.75 from the prior inter-observer agreement analysis were excluded from the training set. A two-stage feature selection was then performed. First, the Minimum Redundancy Maximum Relevance (mRMR) method was applied to the training data to preselect the top 30 informative features. Subsequently, Least Absolute Shrinkage and Selection Operator (LASSO) regression was used for further refinement. The optimal regularization strength (lambda) for the LASSO model was determined via 10-fold cross-validation on the training set, which yielded a final set of 19 non-zero coefficient features for model building.

Three machine learning classifiers—Logistic Regression (LR), SVM, and Random Forest (RF)—were implemented and compared using the selected 19 features. Hyperparameter optimization for each classifier was performed using a grid search strategy coupled with 10-fold cross-validation, again strictly within the training set. Specifically, for LR, the regularization parameter Cwas tuned; for SVM with a radial basis function kernel, both Cand gammawere tuned; for RF, the number of trees (n_estimators, range: 100–500), the maximum tree depth (max_depth, range: 5–20), and the minimum samples per leaf (min_samples_leaf) were optimized. Model selection was based on the average area under the receiver operating characteristic curve (AUC) derived from the inner cross-validation folds of the training set. Subsequently, to statistically validate the superiority of the selected model and formally compare the diagnostic performance of all candidate models, pairwise comparisons of the AUCs on the held-out test set were performed using DeLong’s test.

The model achieving the highest mean cross-validated AUC (the optimized Random Forest) was then retrained on the entirety of the temporary training set (n = 201) using its optimal hyperparameters. This final model was evaluated only once on the completely independent hold-out test set (n = 51) to report its definitive performance metrics, including accuracy, precision, recall, F1-score, and AUC.

Statistical analysis

All analyses were performed using Python (version 3.9) with SciPy (version 1.7.0) and Scikit-learn (version 1.0.0) libraries. Categorical data are presented as frequencies. Model performance is reported with accuracy, precision, recall, F1-score, and AUC with 95% confidence intervals. The calibration of the final model was assessed using calibration curves.Workflow of image post-processing is shown in Fig. 1.

Fig. 1.

Workflow of image post-processing

Results

Baseline characteristics of the study cohort

The comprehensive baseline characteristics of the 252 patients are summarized in Supplementary Table S1. Patients in the fungal sinusitis group (n = 106) were significantly older than those in the chronic sinusitis group (n = 146) (mean age: 54.8 vs. 45.6 years, p < 0.001). Fungal sinusitis presented predominantly unilaterally (84.9% vs. 32.2%, p < 0.001), whereas chronic sinusitis was more frequently bilateral. Regarding conventional CT features, focal hyperdensity (84.9% vs. 6.8%, p < 0.001), bony sclerosis/remodeling (80.2% vs. 56.8%, p < 0.001), and bony erosion (20.8% vs. 6.2%, p < 0.001) were all significantly more common in the fungal sinusitis group, aligning with its known pathophysiological features.

Feature selection and optimization

The rigorous application of the mRMR method initially refined our extensive dataset, isolating 30 features that exhibited the most significant potential for distinguishing between fungal and chronic sinusitis. Subsequent lasso regression analysis meticulously honed this selection down to 19 paramount features, covering 3D shape, first-order intensity, and multi-scale texture (GLCM, GLRLM, GLSZM, NGTDM) and wavelet-derived characteristics. Detailed definitions are provided in Supplementary Table S2. This dual-stage optimization process was instrumental in identifying the most discriminative radiomic signatures, encompassing a diverse array of characteristics including shape, intensity, texture, and model-based features. These selected features emerged as critical indicators, effectively encapsulating the nuanced differences between the two sinusitis conditions (Fig. 2). The distribution differences of each feature between fungal sinusitis and chronic sinusitis are illustrated in the Fig. 3.

Fig. 2.

The feature selection with the mRMR and Lasso: (a) This heatmap displays the Spearman correlation coefficients for the top 30 features selected via the mRMR method. Color intensity indicates the degree of correlation, with red representing positive and blue indicating negative correlations. This visual aid was instrumental in identifying the inter-relationships and selecting the most representative features for further analysis. (b) Depicted here is the lasso regression path, charting the trajectory of feature coefficients against varying alpha values. The optimal alpha, identified by the vertical dashed line, determined the final selection of 19 features that effectively capture the distinct radiomic patterns for the predictive model

Fig. 3.

Distribution of Radiomics Feature Z-Scores by Condition. This bar graph illustrates the Z-scores for selected radiomic features across two conditions: fungal sinusitis and chronic sinusitis. Each feature is represented by a pair of bars, with blue indicating fungal sinusitis and orange denoting chronic sinusitis. The length of the bars reflects the standardized value of each feature, facilitating direct comparison of feature expression between the two conditions

Model performance and evaluation

Upon deploying the logistic regression, SVM, and random forest models, a comprehensive evaluation based on the area under the receiver operating characteristic curve (AUC) was conducted. The random forest model distinguished itself with an exemplary AUC value, signifying its superior diagnostic capability.and plotted the correspondingROC curves with AUC values, as shown in Fig. 4. DeLong’s test revealed that the AUC of the Random Forest model was significantly higher than that of both Logistic Regression (Z = 2.35, p = 0.019) and Support Vector Machine (Z = 2.78, p = 0.005). No significant difference was observed between the AUCs of Logistic Regression and Support Vector Machine (Z = 0.92, p = 0.358).In particular, the optimized random forest model exhibited outstanding performance metrics across the board.

Fig. 4.

Receiver Operating Characteristic (ROC) Curves for Machine Learning Models. This figure presents the ROC curves comparing the performance of three machine learning models: Logistic Regression, Support Vector Machine, and Optimized Random Forest. The AUC indicates the model’s ability to distinguish between classes, with the Optimized Random Forest model achieving the highest AUC of 0.96(0.94–0.98), signifying excellent diagnostic ability

Training Set: Demonstrated an accuracy of 92.51% (88.87% − 96.15%), precision of 92.41% (88.75% − 96.07%), recall of 96.05% (93.36% − 98.74%), and an F1 score of 94.19% (90.96% − 97.42%). The AUC stood at a remarkable 96.04% (94.34% − 97.74%, reflecting the model’s robustness in identifying fungal sinusitis.

Test Set: In an independent validation, the model maintained a high level of accuracy at 91.30% (88.56% − 93.04%), with precision and recall rates of 90.62% (87.62% − 93.62%) and 96.67% (94.65% − 98.70%), respectively, and an F1 score of 93.55% (89.31% − 97.79%). The AUC for the test set was 94.91% (91.31% − 98.41%), underscoring the model’s consistency and reliability in diverse datasets. The detailed scores are shown in Table 1. ROC curve analysis further affirmed the random forest model’s high diagnostic accuracy, (Fig. 5) with an AUC value of 0.95(0.91–0.98) on the test set, showcasing its substantial discriminative power. The confusion matrix provided a clear visualization of the model’s performance (Fig. 6), highlighting its ability to accurately classify fungal sinusitis cases. In the training set, the matrix revealed a high true positive rate, with minimal false negatives, underscoring the model’s sensitivity. Similarly, in the test set, the confusion matrix confirmed the model’s robustness and precision in identifying true cases of fungal sinusitis, with a noteworthy true positive rate and minimal misclassification errors. The calibration curves (Fig. 7), closely hugging the ideal line in both training and test sets, validated the model’s predicted probabilities against the actual outcomes, demonstrating its predictive reliability and accuracy in clinical diagnostics.

Table 1.

Evaluation metrics for the random forest model in training and test sets

Metric	Training Set	Test Set
Accuracy	92.51% (88.87% − 96.15%)	91.30% (88.56% − 93.04%)
Precision	92.41% (88.75% − 96.07%)	90.62% (87.62% − 93.62%)
Recall	96.05% (93.36% − 98.74%)	96.67% (94.65% − 98.70%)
F1 Score	94.19% (90.96% − 97.42%)	93.55% (89.31% − 97.79%)
AUC	96.04% (94.34% − 97.74%)	94.91% (91.31% − 98.41%)

Fig. 5.

ROC Curves for the Optimized Random Forest Model on Training and Test Sets The ROC curves demonstrate the performance of the Optimized Random Forest model on both the training and test datasets. The training set achieved an AUC of 0.96(0.94–0.98), while the test set achieved an AUC of 0.95(0.91–0.98),indicating the model’s consistent and high predictive accuracy across both datasets

Fig. 6.

Confusion Matrices for the Optimized Random Forest Mode. The confusion matrices visually summarize the classification performance of the final model. For the training (development) set (left, n = 201), the model correctly predicted 85 cases as negative (true negatives) and 106 cases as positive (true positives), while incorrectly predicting 7 cases as positive (false positives) and 3 cases as negative (false negatives). This indicates excellent learning capability on the seen data. For the independent test set (right, n = 51), which was held out from all training and development steps, the model correctly identified 24 cases as negative and 23 cases as positive, with only 3 false positives and 1 false negative. The consistent and high performance on this unseen dataset demonstrates the model’s robust generalizability and effective discriminative power

Fig. 7.

Calibration Curves for the Optimized Random Forest Model. The calibration curves in this figure evaluate the probability predictions of the Optimized Random Forest model against actual outcomes for both the training (blue) and test (red) sets. The closer the curves are to the dashed diagonal line, which represents a perfectly calibrated model, the more accurate the predictions. Here, both curves closely approximate this ideal line, with the training set curve almost overlapping it and the test set curve slightly below, indicating a high level of calibration and trustworthiness of the model’s probabilistic outputs

Comparison with conventional CT-based model

To evaluate the incremental value of radiomics beyond routine imaging assessment, a logistic regression model was constructed using four significant conventional CT features: laterality, focal hyperattenuation, bony remodeling/sclerosis, and bony erosion. The conventional CT model yielded a test AUC of 0.782 (95% CI: 0.715–0.840) and accuracy of 76.5% (95% CI: 70.2%–81.9%). In contrast, the radiomics-based random forest model achieved a significantly higher test AUC of 0.949 (95% CI: 0.913–0.984) and accuracy of 91.3% (95% CI: 88.6%–93.0%) (DeLong’s test, p < 0.001). These results confirm that the radiomics model captures subvisual quantitative patterns not detectable by conventional visual evaluation, providing substantial incremental diagnostic value.

Discussion

This study successfully developed and validated a radiomics-based machine learning model for distinguishing fungal sinusitis from chronic sinusitis, demonstrating exceptional diagnostic performance. The optimized random forest model achieved an AUC of 94.91% (91.31% − 98.41%) on the independent test set, highlighting its significant potential for clinical translation. Furthermore, the superiority of the Random Forest model was statistically confirmed, as DeLong’s test demonstrated that its AUC was significantly higher than those of both the Logistic Regression and Support Vector Machine models on the independent test set (both p < 0.05).The radiomics model significantly outperformed the conventional CT-based model, confirming its ability to capture subvisual imaging phenotypes not apparent on routine review. The following discussion will interpret these primary findings in the context of existing literature, elucidate the clinical relevance and mechanistic insights provided by the radiomic features, critically acknowledge the study’s limitations, and propose constructive directions for future research.

Principal findings and model superiority

The strong performance of our model is inherently linked to the biological plausibility of the selected radiomic features. Features related to texture heterogeneity (e.g., GLCM features) likely correspond to the microcalcifications and fungal concretions characteristic of fungal balls, while intensity-based features may reflect the increased density observed in mycetomas. This is consistent with our cohort’s conventional CT findings, where focal hyperdensity and bony changes were significantly more prevalent in the fungal sinusitis group.The Z-score distribution graph visually affirms the significant differences in these feature expressions between the two patient groups, providing a quantitative basis for the model’s decision-making process and strengthening the clinical validity of our approach.

Advancements beyond conventional radiologic assessment

This radiomics-based approach represents a substantial leap beyond traditional CT interpretation. Conventional diagnosis relies on subjective visual assessment of radiographic signs, such as intrasinus hyperdensity, which can be subtle or non-specific [7, 10]. Our model, by contrast, quantifies subvisual textural and morphological patterns within the sinus cavity, translating them into objective, reproducible data. The selection of 19 highly discriminative features through a rigorous process of mRMR and Lasso regression underscores that the model captures complex pathological characteristics invisible to the human eye [11, 12], offering a clear advantage in diagnostic precision and consistency over conventional methods.

Comparative analysis with existing research and novel contributions

When positioned within the emerging field of radiomics for sinonasal diseases, our findings compare favorably with prior efforts. While previous studies have explored radiomics for differentiating sinonasal pathologies [13], few have focused specifically on the critical distinction between fungal and chronic sinusitis with a comparable sample size and methodological rigor. For instance, the model by Zhu et al. [14] targeted eosinophilic chronic rhinosinusitis, a different diagnostic challenge. Our work fills this specific gap. Furthermore, the superior performance of our random forest model, when directly compared to logistic regression and SVM in our study, highlights the value of employing ensemble learning algorithms capable of modeling non-linear relationships within high-dimensional radiomic data, an aspect potentially underexplored in earlier literature [15–17].

Interpretation of high-discriminatory features and pathological basis

The strong performance of our model is inherently linked to the biological plausibility of the selected radiomic features. Features related to texture heterogeneity (e.g., GLCM features) likely correspond to the microcalcifications and fungal concretions characteristic of fungal balls, while intensity-based features may reflect the increased density observed in mycetomas [18, 19]. The Z-score distribution graphvisually affirms the significant differences in these feature expressions between the two patient groups, providing a quantitative basis for the model’s decision-making process and strengthening the clinical validity of our approach.

Clinical implications and workflow integration

The primary clinical value of this model lies in its potential to serve as a non-invasive, rapid decision-support tool. By providing a high-probability objective assessment, it can aid clinicians in prioritizing cases for surgical intervention or initiating pre-operative antifungal therapy, especially in ambiguous scenarios. The high recall rate is particularly valuable for “ruling in” fungal sinusitis with confidence. Integration into existing clinical workflows is feasible, as it operates on standard CT scans, requiring no additional scanning procedures or patient burden [20, 21].

The distinction between fungal sinusitis and chronic sinusitis presents unique diagnostic challenges that differ from those encountered in broader sinonasal disease classifications. Fungal and chronic sinusitis share substantial imaging overlap, including non-specific sinus opacification, mucosal thickening, and sinus wall remodeling. Moreover, fungal balls may coexist with chronic mucosal inflammation, and both conditions can present with unilateral or bilateral involvement. These overlapping features frequently lead to diagnostic uncertainty on conventional CT and often necessitate invasive tissue sampling for definitive diagnosis. This unique challenge underscores the value of radiomics-based approaches that can extract subtle, subvisual imaging features to aid differentiation. Regarding model interpretability, we acknowledge that the ‘black box’ nature of random forest models limits clinical acceptance and trust. Future studies should integrate explainable AI techniques, such as SHAP values, to identify which radiomic features most strongly influence model predictions. Such approaches would not only enhance transparency but may also reveal biologically meaningful imaging biomarkers that could improve understanding of the underlying pathophysiological differences between these two conditions. While SHAP analysis was beyond the scope of the current study, we have identified it as a critical direction for future work to facilitate clinical translation.

Study limitations and future directions

Despite the promising results, several limitations must be acknowledged. Firstly, the model was developed and tested using a single-center cohort, and all imaging data were acquired from scanners of the same vendor with consistent acquisition protocols. No external validation cohort from a different institution or scanner was included. Future multi-center prospective validations with larger, more diverse cohorts are essential to confirm robustness. Secondly, while random forest performed excellently, it acts as a “black box,“ [22] and we did not explore more interpretable AI models or deep learning techniques that might capture even more complex features [23, 24]. Future research should focus on external validation, model interpretability to build clinician trust, and direct comparison with expert radiologists’ performance to definitively establish clinical utility. Finally, the use of ‘CT evidence of increased density within the sinuses’ as an inclusion criterion may have introduced selection bias. This radiological feature is more commonly associated with fungal sinusitis than with chronic sinusitis, and its inclusion may have enriched the fungal sinusitis group with cases that are more readily distinguishable based on imaging alone. Consequently, the diagnostic performance of our radiomics model may be overestimated compared to its expected performance in an unselected patient population presenting with sinonasal opacification.

Conclusion

This study demonstrates the potential of radiomics and machine learning for distinguishing fungal sinusitis from chronic sinusitis. Our random forest model showed promising diagnostic performance using selected radiomic features, offering a potential non-invasive tool to assist clinical decision-making. Further prospective external validation is required before clinical implementation.

Supplementary Information

Below is the link to the electronic supplementary material.

Author contributions

Shaoxiao Li: Data curation, Formal Analysis, Methodology, Software, Writing – original draft. Rujie Li, Qiongfan Ma: Data curation, Software, Writing – original draft. Cheng Zheng, Xiaolin Yan: Formal Analysis, Methodology, Writing – original draft. Shuainan Chen: Conceptualization, Supervision, Validation, Visualization, Writing – review & editing. All authors contributed to the manuscript and approved the final version for submission.

Funding

No Funding.

Data availability

All data generated or analysed during this study are included in this published article.

Declarations

Ethics approval and consent to participate

This paper was approved by Wenzhou hospital of integrated traditional Chinese and western medicine medical ethics Committee (number: 2026-L021), and we were in accordance with the 1975 Helsinki declaration and its later amendments. Given the anonymous and retrospective nature of this study, the Ethics Committee waived the requirement for written informed consent from the participants. All data were processed anonymously to ensure the confidentiality and privacy of the participants.

Consent to participate

Not applicable.

Consent for publication

Not applicable. The manuscript does not include the participant’s identification image or other personal or clinical details.

Generative AI disclosure

No Generative AI was used in the preparation of this manuscript.

Competing interests

The authors declare no competing interests.