Based on a radiomics-clinical nomogram to predict the therapeutic effect of pulmonary cryptococcosis pneumonia Authors: Yan-li Zhang, Chao Ran, Wei Li Published: 2026 ABSTRACT AbstractWe aimed to develop a radiomics-clinical nomogram to predict the therapeutic effect of PC pneumonia. A total of 255 PC pneumonia patients (165 cases with good therapeutic effect and 90 cases with poor therapeutic effect) were retrospectively enrolled from two centers. 190 PC pneumonia patients from Center 1 were randomly divided into a training group (133 cases) and an internal validation group (57 cases) at a ratio of 7:3. The data from Center 2 were used as the external validation group (65 cases). The radiomics features selected from unenhanced CT were input into multiple machine learning models. The optimal model was obtained through receiver operating characteristic (ROC) analysis to construct the radiomics and clinical models. Combined them to build a nomogram for clinical application. These models were evaluated by ROC analysis, calibration curve, and decision curve analysis (DCA). The support vector machine was the best classifier with the highest area under the curve (AUC) of 0.848 in the internal validation group. Although the performance of the nomogram was similar to that of the radiomics model (AUC: 0.886 vs. 0.864; Delong test: P = 0.632), they were significantly higher than that of the clinical model (AUC: 0.886 vs. 0.730, 0.864 vs. 0.730; Delong test, P < 0.05) in the external validation group. The calibration curves and DCAs of the nomogram and radiomics models confirmed their superior predictive agreements and clinical utilities. Radiomics model and nomogram could noninvasively predict the therapeutic effect of PC pneumonia, which was beneficial for precise intervention.Subject terms: Diseases, Medical research Introduction Pulmonary cryptococcosis (PC) is an opportunistic fungal infection. Cryptococcus can activate different inflammatory responses1. The nodule-mass PC is the typical morphological type, characterized by the granuloma2. PC pneumonia has become more common in recent years. Its imaging findings are parenchyma-interstitium opacity due to inflammatory exudation or cryptococcus infiltration3,4. Its clinical symptoms, prognosis, and treatment are more complex than those of nodule-mass PC5. However, the antifungal duration and dosage are inclusive6. Antifungal drugs have some adverse reactions. Unreasonable antifungal administration is harmful to PC patients7. The antifungal principle is to maximize therapeutic efficacy and minimize side effects8. The balance between antifungal efficacy and adverse reaction is a challenge in PC treatment9–11. The antifungal regimen should be adjusted according to the therapeutic response. For PC pneumonia with good therapeutic effect, the antifungal dosage and duration should be decreased9–11. For PC pneumonia with poor therapeutic effect, antifungal treatment should be intensified9–11. Therefore, it is important to predict the therapeutic effect of PC pneumonia. Nonspecific biomarkers and incomplete follow-up weakened the antifungal evaluation. The antifungal influencing factors may lie in the initial imaging and clinical data. Microscopic heterogeneity combined with clinical information can aid medical decision-making. At present, radiomics research mainly focuses on differentiation between nodular-mass PC and solitary pulmonary tumors12,13. However, radiomics performed well in therapeutic prediction of other pneumonia14,15. Combining clinical variables with artificial neural networks can effectively predict in-hospital mortality and mechanical ventilation risk in pulmonary inflammation16,17. Therefore, this study will apply the clinical-radiomic nomogram to predict the therapeutic effect and provide personalized regimens for patients with PC pneumonia. Materials and methods The acquisition and analysis of clinical and CT data This study was designed and conducted following the Declaration of Helsinki and approved by the Ethics Committees of Affiliated Yantai Yuhuangding Hospital of Qingdao University (No. 2023 − 275) and Affiliated Hospital of Yangzhou University (No. 2023-YKL07). Since the retrospective study, the patients’ handwritten informed consent forms were waived. From June 2015 to July 2024, clinical and imaging data from 354 patients with PC pneumonia were consecutively collected in two independent centers. All patients underwent unenhanced chest CT (16 or 64-section CT scanner, GE Healthcare, Milwaukee, USA, tube voltage = 120 kV, tube current = 10–300 mA, collimation width = 0.625 mm×128–0.625 mm×32, spiral pitch = 1,0, slice thickness = 5 mm, slice interval = 5 mm, matrix = 512 × 512) before treatment. The 1.0 mm thickness was used as a reconstruction standard and applied to all CT images. After CT acquisition, ComBat correction was performed to eliminate the multi-center imaging differences18,19. Based on CT imaging, PC pneumonia was defined as patchy consolidation or interstitial opacity. Inclusion criteria: definite pathological diagnosis, positive cryptococcal antigen (CrAg, colloidal gold labeled) or etiological culture, sufficient imaging quality, and standardized antifungal therapy. Exclusion criteria: incomplete clinical and imaging data (30 cases), with other pulmonary lesions (26 cases), infection with other pathogens (23 cases), and antimicrobial therapy before CT scan (20 cases). Combined treatment was administered to patients with PC pneumonia, including induction therapy (amphotericin B, 0.5-1.0 mg/kg/day and 5-fluorocytosine, 100 mg/kg/day, at least 4 weeks), consolidation therapy (fluconazole, 400 mg/day, 8 weeks), and maintenance therapy (fluconazole, 200 mg/day, 6–12 months)9–11. Based on the antifungal process, therapeutic evaluations were conducted before the completion of induction treatment (4 weeks) and consolidation treatment (8 weeks) to determine the next antifungal plan. During maintenance therapy, the follow-up should be conducted at 6 to 12 months. If the condition changed, the follow-up schedule would be adjusted accordingly. Both centers followed the standardized antifungal regimens and treatment durations to ensure the therapeutic response. All patients in this study were divided into two groups based on therapeutic effects: PC pneumonia with good therapeutic effect and PC pneumonia with poor therapeutic effect. The good therapeutic effect was defined as improved clinical symptoms, imaging absorption, negative CrAg, and without recurrence or dissemination during antifungal treatment. Conversely, it indicated the poor therapeutic effect. According to the inclusion and exclusion criteria, 255 patients with PC pneumonia were consecutively recruited from Center 1 (190 cases, including 125 cases with good therapeutic effect and 65 cases with poor therapeutic effect) and Center 2 (65 cases, including 40 cases with good therapeutic effect and 25 cases with poor therapeutic effect). Ultimately, 190 patients with PC from Center 1 were randomly divided into a training group (133 cases) and an internal validation group (57 cases) at a ratio of 7:3 through random division. The selections of clinical features, radiomic features, and predictive models were based on the training and internal validation groups. The data from Center 2 were used as the external validation group (65 cases). Clinical indicators included sex, age, environmental exposure, respiratory symptoms, inflammatory response, bilateral distribution, CT morphology (consolidation, consolidation-node or consolidation-interstitium), halo sign, pleural-based involvement, pleural effusions, and mediastinal lymphadenopathy. The inflammatory response met one or more of the following indicators: white blood cell count > 10 × 109/L, neutrophil percentage > 70%, and C-reactive protein > 8.2 mg/L20,21. Consolidation was defined as the patchy opacity obscuring the underlying structures. The ground-glass reticular opacity without obscuring the underlying structure was defined as the interstitial pattern. The ground-glass opacity around the lesion was referred to as a halo sign. Pleural involvement was defined as the focal or extensive pleural attachment of the lesion. A lymph node with a short axis diameter of > 10 mm was defined as lymphadenopathy. These clinical and imaging indicators were evaluated by an infectiologist with 10 years of experience and a radiologist with 10 years of experience without knowing the antifungal effects. In case of disagreement, a consensus was reached through negotiation. CT data preprocessing, lesion segmentation, and feature extraction Based on the Onekey AI platform (v.2.2.3, https://www.medai.icu), OKT-toolkits were used for image resampling (1 mm×1 mm×1 mm) and normalization to reduce the influence of different CT scanners and resolutions. A radiologist with five years of experience (reader 1) used ITK-SNAP to delineate the lesions on lung window CT (window width 1000–2000 Hu; window level − 800 - -450 Hu) layer by layer. Based on segmentation proficiency and lesion complexity, the segmentation time of each case was 3 to 5 min using semi-automatic ITK-SNAP. Observed the segmentation effect and made manual corrections. Furthermore, 20% of all cases were randomly selected to be delineated by another radiologist with ten years of experience (reader 2) for inter-observer interclass correlation coefficients (ICC) evaluation22–25. The inconsistent segmentation was finally determined by a radiologist with twenty years of experience (reader 3). Two weeks later, 20% of all cases were randomly selected to be delineated by reader 1 for intra-observer ICC evaluation. ICC > 0.75 indicated good consistency. The region of interest (ROI) should cover the entire lesion, avoiding adjacent normal tissues. For multiple lesions, the largest one was delineated26,27. To extract radiomics features from ROIs and perform z-score normalization on them through Pyradiomics. Feature scaling was performed separately within training and validation groups to avoid data leakage. Radiomics features selection and radiomics signature development The dimensionality reduction and radiomics features selection were as follows. First, all these radiomics features were screened through univariate analysis (Mann-Whitney U-test or independent T-test) to retain those significant features (P < 0.05). Second, Spearman’s rank correlation coefficient was calculated to exclude strongly correlated features, selecting one of the two features with a correlation coefficient > 0.9. To reduce the feature redundancy by a greedy recursive deletion. Third, feature selection was conducted by the least absolute shrinkage and selection operator (LASSO) regression. All irrelevant feature coefficients are adjusted to zero by the weight of λ. The 10-fold cross-validation was used to find the optimal λ. The LASSO regression modeling was performed by the Python scikit-learn package. Six machine learning models were used in this study, including random forest (RF), support vector machine (SVM), extreme gradient boosting (XGBoost), decision tree (DT), light gradient boosting machine (LightGBM), and Extra Trees (ET). The selected radiomics features were applied to these machine learning models, and the model with the highest area under the curve (AUC) was selected to construct the radiomics signature. The parameters used by the best model were selected as the optimal hyperparameters in this study. Our parameter tuning was based on 5-fold cross-validation and grid search. All possible combinations of hyperparameters were used for model training, parameter tuning, and overfitting prevention in the training and internal validation groups. Evaluated the final model generalization in the external validation group. The development of a clinical signature and a nomogram The clinical signature building was similar to the radiomics signature. Based on the clinical baseline statistics, the features with P < 0.05 in the training group were selected. For the objective comparison with the radiomics signature, the same machine learning model and fixed cohort were used to build the clinical signature. Combining radiomics and clinical signatures to construct a nomogram for clinical application. The three models were tested in the internal and external validation groups. The methodology flowchart of this study is shown in Fig. 1. The methodology flowchart of this study. Statistical analysis In this study, statistical analysis was conducted using SPSS 26 (IBM Corp., Armonk, NY, USA) and the Onekey AI platform (v.2.2.3, https://www.medai.icu). For the continuous variables, the Shapiro-Wilk normality test and the Levene homoscedasticity test were performed before the comparative analyses. All data were presented as numbers (percentage) or mean ± SD. The continuous variables were compared by the independent T-test (two-tailed) or the Mann-Whitney U-test. The statistical analyses of dichotomous variables were conducted by the Chi-square test or Fisher’s exact test (two-tailed). The predictive efficacy of different models was evaluated by the receiver operating characteristic (ROC) curve analysis and the DeLong test. The model predictions and actual results were analyzed by the calibration curve and the Hosmer-Lemeshow test. The decision curve analysis (DCA) was used to assess the clinical utility of the three models by the net benefit at different probability thresholds. P < 0.05 was defined as statistically significant Table 1 and 2. Table 1Baseline analyses of patients with PC pneumonia in different groups. Values are given as n, n (%) or mean ± SD. PC pulmonary cryptococcosis, CT computed tomography. Significant values are in bold.Baseline featuresTraining group(n = 133)Internal validation group(n = 57)External validation group(n = 65) P Sex (male/female)67/6627/3035/300.773Age (years)55.1 ± 6.954.2 ± 7.252.9 ± 6.60.419Environmental exposure24 (18.0%)15 (26.3%)14 (21.5%)0.430Respiratory symptom96 (72.2%)38 (66.7%)50 (76.9%)0.452Inflammatory response50 (37.6%)23 (40.4%)28 (43.1%)0.754Immunosuppression30 (22.6%)16 (28.1%)20 (30.8%)0.424Halo sign60 (45.1%)29 (50.9%)35 (53.8%)0.477Bilateral distribution55 (41.4%)28 (49.1%)30 (46.2%)0.578Consolidation42 (31.6%)19 (33.3%)23 (35.4%)0.864Consolidation-node59 (44.4%)24 (42.1%)27 (41.5%)0.917Consolidation-interstitium32 (24.1%)14 (24.6%)15 (23.1%)0.980Pleural-based involvement99 (74.4%)39 (68.4%)51 (78.5%)0.447Pleural effusions18 (13.5%)9 (15.8%)8 (12.3%)0.852Mediastinal lymphadenopathy23 (17.3%)11 (19.2%)14 (21.5%)0.769 Baseline analyses of patients with PC pneumonia in different groups. Values are given as n, n (%) or mean ± SD. PC pulmonary cryptococcosis, CT computed tomography. Significant values are in bold. Table 2Baseline comparison of PC pneumonia with different therapeutic effects in the training group. PC pulmonary cryptococcosis. Values are given as n, n (%) or mean ± SD. Significant values are in bold.Baseline featuresPC pneumonia with good therapeutic effect(n = 87)PC pneumonia withpoor therapeutic effect(n = 46) P Sex (male/female)46/4121/250.428Age (years)53.8 ± 6.155.1 ± 8.00.278Environmental exposure16 (18.4%)8 (17.4%)0.887Respiratory symptom60 (69.0%)36 (78.3%)0.255Inflammatory response18 (20.7%)32 (69.6%) < 0.001 Immunosuppression9 (10.3%)21 (45.7%) < 0.001 Halo sign42 (48.3%)18 (39.1%)0.313Bilateral distribution20 (23.0%)35 (76.1%) < 0.001 Consolidation32 (36.8%)10 (21.7%)0.076Consolidation-nodular pattern43 (49.4%)16 (34.8%)0.106Consolidation-interstitial pattern12 (13.8%)20 (43.5%) < 0.001 Pleural-based involvement62 (71.3%)37 (80.4%)0.249Pleural effusions10 (11.5%)8 (17.4%)0.344Mediastinal lymphadenopathy13 (14.9%)10 (21.7%)0.296 Baseline comparison of PC pneumonia with different therapeutic effects in the training group. PC pulmonary cryptococcosis. Values are given as n, n (%) or mean ± SD. Significant values are in bold. Results Radiomics features and performance of the radiomics model The consistency and repeatability of ROI segmentation were good (intra-observer ICC = 0.94, inter-observer ICC = 0.92) in this study. A total of 1008 radiomics features were extracted from the ROIs, including first-order features (230), shape features (18), and texture features (760). According to different extraction methods, texture features could be further divided into the gray-level co-occurrence matrix (GLCM, 233), gray-level run length matrix (GLRLM, 170), gray-level size zone matrix (GLSZM, 166), gray-level dependence matrix (GLDM, 126), and neighborhood gray-tone difference matrix (NGTDM, 65) features. After dimensionality reduction, six non-zero coefficient radiomics features were selected and integrated into the radiomics signature, including 4 wavelet features and 2 Laplacian of Gaussian (LoG) filtering features. Among the six machine learning models, the SVM performed best in the internal validation group (AUC = 0.848, 95% CI = 0.734–0.963, accuracy = 0.833, sensitivity = 0.733, specificity = 0.879). Therefore, SVM was selected as the base model to build the radiomics and clinical signatures. The radiomics signature performed well in training group (AUC = 0.947, 95% CI = 0.906–0.989, accuracy = 0.898, sensitivity = 0.891, specificity = 0.901), internal validation group (AUC = 0.860, 95% CI = 0.731–0.988, accuracy = 0.809, sensitivity = 0.893, specificity = 0.766) and external validation group (AUC = 0.864, 95% CI = 0.735–0.993, accuracy = 0.812, sensitivity = 0.889, specificity = 0.771). See Figs. 1a-c and 2a-c, and Tables 3 and 4 for details. The dimensionality reduction and radiomic feature selection based on LASSO. For the discrimination between PC pneumonia with good effect and PC pneumonia with poor effect, the distribution of LASSO coefficients with different log (λ) was given (a), and the optimal parameter λ was selected through 10-fold cross-validated MSE (b). The histogram of the coefficients was based on the selected features (c). LASSO least absolute shrinkage selection operator, MSE mean standard error. Table 3Predictive performance of six machine-learning models in the training and internal validation groups. SVM support vector machine, DT decision tree, RF random forest, XGBoost Extreme Gradient Boosting, ET Extra Trees, LightGBM light gradient boosting machine, AUC area under the curve, CI confidence interval.ModelsGroupAUC (95% CI)AccuracySensitivitySpecificitySVMTraining0.856 (0.793–0.919)0.7620.8670.713Internal validation0.848 (0.734–0.963)0.8330.7330.879DTTraining1.000 (nan - nan)1.0001.0001.000Internal validation0.696 (0.551–0.842)0.7080.6671.000RFTraining0.999 (0.996–1.000)0.9891.0000.984Internal validation0.807 (0.675–0.939)0.7710.8000.787XGBoostTraining1.000 (nan - nan)1.0001.0001.000Internal validation0.810 (0.666–0.954)0.8330.6670.909ETTraining1.000 (nan - nan)1.0001.0001.000Internal validation0.797 (0.662–0.932)0.7080.8590.636LightGBMTraining0.918 (0.879–0.955)0.8310.9000.798Internal validation0.784 (0.629–0.939)0.8130.6000.909 Predictive performance of six machine-learning models in the training and internal validation groups. SVM support vector machine, DT decision tree, RF random forest, XGBoost Extreme Gradient Boosting, ET Extra Trees, LightGBM light gradient boosting machine, AUC area under the curve, CI confidence interval. Table 4Predictive performances and comparisons of three models in the training and validation groups. AUC area under the curve, CI confidence interval, Rad radiomics. Significant values are in bold.ModelsGroupAUC(95% CI)AccuracySensitivitySpecificityAUC vs.RadAUC vs. NomogramClinic SignatureTraining0.707(0.617–0.797)0.6820.5810.743P < 0.001P < 0.001Internal validation0.656(0.459–0.852)0.6390.6040.656P < 0.001P < 0.001External validation0.730(0.546–0.915)0.7070.6260.766P = 0.002P = 0.001Rad signatureTraining0.947(0.906–0.989)0.8980.8910.901-P = 0.879Internal validation0.860(0.731–0.988)0.8090.8930.766-P = 0.883External validation0.864(0.735–0.993)0.8120.8890.771-P = 0.632NomogramTraining0.951(0.909–0.992)0.9020.8890.912P = 0.879-Internal validation0.864(0.736–0.992)0.8730.8490.856P = 0.883-External validation0.886(0.766–1.000)0.8690.8360.903P = 0.632- Predictive performances and comparisons of three models in the training and validation groups. AUC area under the curve, CI confidence interval, Rad radiomics. Significant values are in bold. 0.707 (0.617–0.797) 0.656 (0.459–0.852) 0.730 (0.546–0.915) 0.947 (0.906–0.989) 0.860 (0.731–0.988) 0.864 (0.735–0.993) 0.951 (0.909–0.992) 0.864 (0.736–0.992) 0.886 (0.766–1.000) Clinical features and performance of the clinical model In this study, patients with PC pneumonia had some severe respiratory symptoms, including cough, expectoration, chest distress, chest pain, and fever. There were no differences in clinical baseline data among the training, internal, and external validation groups, meaning a reasonable data division. Compared to PC pneumonia with good therapeutic effect, PC pneumonia with poor therapeutic effect was more common with inflammatory response (20.7% vs. 69.6%, P < 0.001), immunosuppression (10.3% vs. 45.7%, P < 0.001), bilateral distribution (23.0% vs. 76.1%, P < 0.001), and consolidation-interstitium pattern (13.8% vs. 43.5%, P < 0.003) in the training group. There were no differences in sex, age, environmental exposure, respiratory symptoms, halo sign, consolidation, consolidation-node pattern, pleural-based involvement, pleural effusions, and mediastinal lymphadenopathy between the two groups. Therefore, immunosuppression, inflammatory response, bilateral distribution, and consolidation-interstitium pattern were selected to build a clinical signature. The efficacy of clinical signature was not good in the training group (AUC = 0.707, 95% CI = 0.617–0.797, accuracy = 0.682, sensitivity = 0.581, specificity = 0.743), internal validation group (AUC = 0.656, 95% CI = 0.459–0.852, accuracy = 0.639, sensitivity = 0.604, specificity = 0.656) and external validation group (AUC = 0.730, 95% CI = 0.546–0.915, accuracy = 0.707, sensitivity = 0.626, specificity = 0.766). See Tables 2 and 4, and Figs. 4a-c for details. ROC curves of predictive models for PC pneumonia with different therapeutic effects. Compared with the clinical signature, the nomogram and radiomics signature showed similar and better performance in the training group (a), internal validation group (b), and external validation group (c). ROC receiver operating characteristic. Performance of the nomogram Combining radiomics and clinical signatures, the nomogram showed good performance in the training group (AUC = 0.951, 95% CI = 0.909–0.992, accuracy = 0.902, sensitivity = 0.889, specificity = 0.912), internal validation group (AUC = 0.864, 95% CI = 0.736–0.992, accuracy = 0.873, sensitivity = 0.849, specificity = 0.856) and external validation group (AUC = 0.886, 95% CI = 0.766–1.000, accuracy = 0.869, sensitivity = 0.836, specificity = 0.903). Delong tests showed that the performances of the nomogram and radiomics models were significantly higher than that of the clinical model in the training group (AUC: 0.951 vs. 0.707, 0.947 vs. 0.707; P < 0.05, respectively), internal validation group(AUC: 0.864 vs. 0.656, 0.860 vs. 0.656; P < 0.05, respectively) and external validation group (AUC: 0.886 vs. 0.730, 0.864 vs. 0.730; P < 0.05, respectively). There was no predictive difference between the nomogram and radiomics model in the training group (AUC, 0.951 vs. 0.947, P = 0.879), internal validation group (AUC, 0.864 vs. 0.860, P = 0.883), and external validation group (AUC, 0.886 vs. 0.864, P = 0.632). The confidence intervals for calibration and decision curves were based on 2000 Bootstrap resamples. The calibration curves of clinical, radiomics, and nomogram models showed good agreement between actual and predicted labels in the internal and external validation groups, without statistical differences in the Hosmer-Lemeshow tests (P > 0.05). In the internal validation group, the slopes of clinical model, radiomics model, and nomogram were 1.08 (95% CI 0.96–1.13), 1.06 (95% CI 0.93–1.11), and 1.09 (95% CI 0.94–1.14), respectively; the intercepts of clinical model, radiomics model, and nomogram were − 0.07 (95% CI -0.11–0.09), -0.05 (95% CI -0.08–0.06), and − 0.08 (95% CI -0.13–0.08), respectively. In the external validation group, the slopes of clinical model, radiomics model, and nomogram were 1.04 (95% CI 0.98–1.08), 1.06 (95% CI 0.95–1.10), and 1.07 (95% CI 0.93–1.11), respectively; the intercepts of clinical model, radiomics model, and nomogram were − 0.02 (95% CI -0.06–0.03), -0.04 (95% CI -0.09–0.05), and − 0.05 (95% CI -0.10–0.07), respectively. Compared with the clinical model, DCAs showed that the nomogram and radiomics models could bring about more clinical benefits in internal and external validation groups. In the internal validation group, within the threshold probability range of 0.2 to 0.4, the net benefit of clinical model was 0.14 (95% CI 0.09–0.21); within the threshold probability range of 0.1 to 0.8, the net benefit of radiomics model was 0.22 (95% CI 0.15–0.30); within the threshold probability range of 0.2 to 0.4, the net benefit of nomogram was 0.23 (95% CI 0.16–0.31). In the external validation group, within the threshold probability range of 0.3 to 0.6, the net benefit of clinical model was 0.08 (95% CI 0.02–0.17); within the threshold probability range of 0.1 to 0.9, the net benefit of radiomics model was 0.25 (95% CI 0.18–0.33); within the threshold probability range of 0.3 to 0.6, the net benefit of nomogram was 0.28 (95% CI 0.19–0.35). See Table 4; Figs. 3 and 4a-c, and Figs. 5a-d for details. Clinical application of the nomogram in the prediction of PC pneumonia with different therapeutic effects. By integrating clinical and radiomics signatures, the different antifungal effects could be inferred. Calibration curves and decision curves of different models. The clinical, radiomics, and nomogram models showed good consistency between predicted and actual labels in the internal validation group (a) and external validation group (b). The decision curves of the nomogram and radiomics models were above the two reference lines, indicating their clinical benefits. Compared with the clinical model, the nomogram and radiomics models brought about more benefits over a larger threshold range in the internal validation group (c) and external validation group (d). Discussion The antifungal influencing factors have complex interactions. The structural heterogeneity is the foundation of radiomics28. The different gray-scale levels and spatial distribution of PC pneumonia can be quantified by radiomics features. Because the radiomics signature outperformed the clinical signature, the nomogram displayed good performance and clinical applicability. Our results indicated that immune status, inflammatory response, lesion distribution, and CT morphology were related to the therapeutic effect of PC pneumonia. Immunocompromised PC was more common with bilateral parenchyma-interstitium involvement, reflecting its obvious invasiveness29. Without immune protection, PC invaded the pulmonary parenchyma and interstitium30. The extensive distribution of lesions caused significant inflammatory reactions, clinical severity, and antifungal difficulty. The immunocompetent PC lesions were mostly unilateral solid exudation, with few interstitial changes. Effective immune defense limited the PC infiltration31. Normal immune function, unilateral distribution, and localized exudation reduced inflammatory response and enhanced antifungal efficacy. Unfortunately, non-specific subjective clinical features could not improve predictive performance. The value of clinical variables in this study was limited due to the clinical complexity of PC pneumonia. The identification of the semantic or subjective features was individualized. These non-negligible variabilities could account for the inefficiency of the clinical model. Radiomics can explore high-throughput microscopic features and provide an objective antifungal prediction of PC pneumonia32. Although only six radiomics features were selected for antifungal prediction in this study, they were all wavelet and LoG features with good discriminative capabilities. Wavelet features are multi-scale detail information extracted through the wavelet transform33. The wavelet transform decomposes the image into low-frequency and high-frequency components, displaying the main outline and local details of PC pneumonia34. The LoG features are extracted from the original grayscale values through LoG filtering, highlighting the imaging contrast and details35. The wavelet transform and LoG filtering of radiomics features improved the predictive accuracy and stability36. Furthermore, wavelet and LoG features could control the interference of noise and segmentation errors37. These spatial heterogeneities ensured the therapeutic assessment of PC pneumonia. SVM can solve the overfitting tendency of small sample data, showing good generalization and interpretability38. Although the value of clinical variables was poor, the efficacy of the nomogram was similar to that of the radiomics SVM model. This reflected the robustness of radiomic features from another perspective. When the radiomics features encompass the key information, the gain of the combined model may not be significant23. Based on the robust radiomics features, clinical variables with limited value did not reduce the performance of the nomogram. In this study, the radiomic signature and clinical-radiomic nomogram showed more benefits than the clinical signature in antifungal prediction. Because the therapeutic response of PC pneumonia was complicated, it was difficult to evaluate the antifungal effect subjectively39. Due to the retrospective nature, selection bias cannot be avoided. Although this study was a double-center retrospective study, the small external validation group restricted the generalizability of our results. Subsequent large-sample prospective studies are necessary. Indeed, univariate selection of radiomics feature only evaluates the independent association between a variable and outcome, ignoring interactions and non-linear relationships among variables. Furthermore, the development of multi-omics is the future. Conclusion Radiomics features, inflammatory response, immunosuppression, bilateral distribution, and consolidation-interstitium pattern affected the structural heterogeneity and antifungal treatment of PC pneumonia. The radiomics revealed the correlation between microscopic features and therapeutic effect. Radiomics signature and nomogram could noninvasively predict the antifungal effect of PC pneumonia, making up for the deficiencies of conventional clinical-imaging information. This was crucial for stratified management and medical decision-making of PC pneumonia. Author contributions Y.L.Z. conducted clinical research and manuscript preparation. C.R. and Y.L.Z. conducted data collection and interpretation of results. W.L. and C.R. conducted imaging analysis. W.L. conducted design and supervision of the study. Funding This research did not receive any specific grant from funding agencies in the public, commercial, or not-for-profit sectors. Data availability Data in this study are available from the corresponding author upon reasonable request. Declarations Competing interests The authors declare no competing interests. Footnotes Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. Yan-li Zhang and Chao Ran contributed equally to this work. References Associated Data Data Availability Statement Data in this study are available from the corresponding author upon reasonable request.