Machine-learning model for differentiating round pneumonia and primary lung cancer using CT-based radiomic analysis

Abstract

Background:

Round pneumonia is a benign lung condition that can radiologically mimic primary lung cancer, making diagnosis challenging. Accurately distinguishing between these diseases is critical to avoid unnecessary invasive procedures. This study aims to distinguish round pneumonia from primary lung cancer by developing machine-learning models based on radiomic features extracted from computed tomography (CT) images.

Methods:

This retrospective observational study included 24 patients diagnosed with round pneumonia and 24 with histopathologically confirmed primary lung cancer. The lesions were manually segmented on the CT images by 2 radiologists. In total, 107 radiomic features were extracted from each case. Feature selection was performed using an information-gain algorithm to identify the 5 most relevant features. Seven machine-learning classifiers (Naïve Bayes, support vector machine, Random Forest, Decision Tree, Neural Network, Logistic Regression, and k-NN) were trained and validated. The model performance was evaluated using AUC, classification accuracy, sensitivity, and specificity.

Results:

The Naïve Bayes, support vector machine, and Random Forest models achieved perfect classification performance on the entire dataset (AUC = 1.000). After feature selection, the Naïve Bayes model maintained a high performance with an AUC of 1.000, accuracy of 0.979, sensitivity of 0.958, and specificity of 1.000.

Conclusion:

Machine-learning models using CT-based radiomics features can effectively differentiate round pneumonia from primary lung cancer. These models offer a promising noninvasive tool to aid in radiological diagnosis and reduce diagnostic uncertainty.

1. Introduction

The incidence and mortality of lung cancer have rapidly increased since the 1930s, largely owing to the increasing prevalence of smoking.^[1] Smoking is the most significant risk factor for lung cancer; however, radon and asbestos exposure also play a role in its pathogenesis.^[2] Lung cancer remains the leading cause of cancer-related deaths associated with smoking (accounting for 80%–85% of such deaths). Globally, it is the most common cancer type in men, and the 4th most common cancer in women.^[3] The 5-year survival rate is approximately 16%.^[2]

Round pneumonia, a rare type of lobar pneumonia in adults, typically arises from congenital abnormalities in connective tissue structures such as the pores of Kohn and canals of Lambert. These structural anomalies confine infection to a single lung lobe. Streptococcus pneumoniae is often the causative agent, and on imaging, round pneumonia usually appears as a well-circumscribed, round area of consolidation that resembles a lung mass. Consequently, imaging features of round pneumonia can mimic those of malignant lung masses. In clinical practice, lung lesions are primarily detected by computed tomography (CT), which provides high contrast resolution and rapid image acquisition. Moreover, several studies have shown that a single tumor biopsy may not adequately characterize tumor morphology owing to intratumoral heterogeneity.^[4] Lung biopsy is an invasive procedure with a risk of complications such as pneumothorax. Consequently, many researchers have aimed to differentiate round pneumonia from malignant lung lesions based solely on radiological findings. Various diagnostic strategies have been employed, including histopathological evaluation, radiomic analyses, and machine-learning methodologies.^[5–7]

Radiomics is a technique that extracts quantitative data from medical imaging modalities, such as CT, positron emission tomography, MRI, and ultrasound. By evaluating features, such as shape, size, texture, and attenuation, radiomics transforms visual attributes, which are too subtle or numerous for manual interpretation by radiologists, into structured datasets. Radiomics can be likened to taking a photograph of a fruit and analyzing the image to create a detailed profile of the fruit size, shape, surface texture, color, and other physical properties. These detailed data can then be analyzed using statistical or artificial intelligence techniques to predict with high accuracy that the object is a fruit (and not a car) and, more specifically, to determine whether it is a Granny Smith apple (as opposed to another variety) and even estimate its degree of ripeness. The fundamental goal of radiomics is to ensure that phenotypic data obtained from images reflect the underlying biological behavior and genetic characteristics of lesions. This approach of linking imaging phenotypes to genotypes is referred to as “radiogenomics.”^[8]

Lung imaging presents a unique challenge for both radiologists and radiomics systems because many benign and malignant processes can appear qualitatively similar on scans. Over time, radiologists have identified a relatively small number of qualitative imaging features to distinguish benign from malignant lesions, such as speculation, lesion size, attenuation, and the presence of cystic air spaces around the lesion.^[9] The current challenge in thoracic radiomics is 2-fold: first, to develop systems that can accurately extract phenotypic features from images; second, to identify potentially thousands of features that are associated with the underlying genotype and disease behavior in order to support prognosis and clinical management.^[10–12]

To the best of our knowledge, no prior study has specifically used CT-derived radiomic analysis to differentiate round pneumonia from malignant lung masses. In this context, our research offers a novel strategy for distinguishing between these 2 radiologically similar conditions. The primary aim of this study was to develop machine-learning models based on radiomic features extracted from CT images to effectively differentiate round pneumonia from primary lung cancer, and to compare the diagnostic performance of these models.

2. Materials and methods

This study was approved by the Institutional Ethics Committee of Elaziğ Fethi Sekin City Hospital (Radiology Department) (Approval Date: May 8, 2025; Approval No: 2025/9-21). Due to the retrospective nature of the study, the requirement for written informed consent was waived by the committee. Patients diagnosed with round pneumonia or non-small cell lung cancer (NSCLC) between 2020 and 2025 were identified from the hospital database. All thoracic CT scans were performed using a 128-slice CT scanner (Ingenuity Core 128; Philips Medical Systems, Best, Netherlands). Axial images were obtained craniocaudally from the thoracic inlet to the midkidneys. CT scans were performed without intravenous contrast using the following parameters: 100 kVp, 100 to 170 mA (with automatic tube current modulation), 2 × 2 mm collimation, 0.5 mm slice thickness, and 1 mm isotropic voxel size. The images were reviewed by 2 radiologists (H.G. and M.Y.). The inclusion criteria were as follows: patients diagnosed at our institution between 2020 and 2025 with round pneumonia or a lung mass detected on CT; at least 1 high-quality thoracic CT scan digitally archived in the hospital PACS; age 18 years or older; and complete clinical and radiological data available. In the round pneumonia group, patients showed clinical and radiological improvements after antibiotic therapy, with regression of the lesion on follow-up imaging. The lung cancer group included patients histopathologically diagnosed with NSCLC, specifically adenocarcinoma (n = 12), “bronchial carcinoma” (n = 7) and squamous cell carcinoma (n = 5).

Exclusion criteria included low-quality or artifact-prone CT images; lack of definitive histopathological diagnosis; incomplete clinical or treatment records; a history of extrapulmonary malignancy of uncertain relation to the pulmonary lesion; noninfectious etiologies such as granulomatous disease, vasculitis, or metastases; patients under 18 years of age; and the presence of multiple pulmonary lesions in the same patient that could not be distinguished from each other. Ultimately, 48 patients meeting the inclusion criteria (24 with round pneumonia and 24 with primary lung cancer) were included in the study (Fig. 1A and B). This sample size met the minimum statistical requirement and supported the validity and robustness of the analytical results.

Figure 1.

(A) Malignant mass adjacent to the pleura in the right lung (thick black arrow); (B) Broad-based round pneumonia lesion in the left lung resembling a mass (thin black arrow).

All thoracic CT images used in this study were retrieved from the institution’s PACS, exported in JPEG format, and processed using the 3D Slicer software (version 4.10.2). Image preprocessing steps included contrast enhancement, ±3 sigma normalization, N4ITK bias field correction, and isotropic resampling to a voxel size of 1 mm × 1 mm × 1 mm. Lesion segmentation was performed manually by 2 experienced chest radiologists (H.G. and M.Y.), using axial slices at the level of the longest diameter of the lesion. Spiculated extensions and peritumoral fibrotic streaks were excluded from segmentation (Fig. 2). Representative images of lesions were selected to illustrate the findings; no artificial modifications were made to the images. Radiomic features were then extracted from each lesion and the resulting dataset was used to train the machine-learning algorithms. These image-based analyses play an important role in accurately distinguishing between benign and malignant lesions and in developing high-performance AI models.

Figure 2.

Manual segmentation of a lung tumor using 3D Slicer software.

Using the radiomics module of the same software, 107 radiomic features were extracted from segmented lesions. These comprised 14 shape-based semantic descriptors, 18 first-order (histogram) features, 24 second-order texture features derived from the Gray-Level Co-occurrence Matrix (GLCM), 51 higher-order texture features derived from advanced matrices, such as the Neighborhood Gray Tone Difference Matrix, Gray-Level Run Length Matrix, Gray-Level Size Zone Matrix, and Gray-Level Dependence Matrix.

2.1. Machine-learning model development

Machine-learning models were developed using the Orange Data Mining software (version 3.24; University of Ljubljana, Faculty of Computer and Information Science, Ljubljana, Slovenia; https://orangedatamining.com). The algorithms used included a Support Vector Machine (SVM), Logistic Regression, k-nearest neighbors (k-NN), Naïve Bayes, Decision Trees, Random Forest, and Neural Networks. All models were trained with default hyperparameter settings; this choice was deliberately made to limit the model complexity and enhance generalizability. Additionally, algorithms such as SVM and Random Forest have intrinsic regularization mechanisms, which provide natural resistance to overfitting.

To evaluate model performance, classification metrics such as area under the ROC curve (AUC), accuracy, sensitivity, and specificity were used. The performance assessment was initially performed using 10-fold stratified cross-validation. However, to objectively evaluate the impact of hyperparameter tuning on model performance and minimize overfitting risk, the modeling process was carried out using a nested cross-validation protocol. In this protocol, the dataset was split in an outer loop into training and test sets using 10-fold stratified cross-validation. Within each outer fold, only the training data were used in an inner loop of 5-fold cross-validation for hyperparameter optimization. Thus, the test data remained completely isolated, allowing for an unbiased evaluation of the model’s generalizable performance.

The feature-selection process was structured to prevent any risk of data leakage. In each outer cross-validation fold, the information-gain algorithm was applied only to the training subset to select the 5 most informative features and the model was trained on these features. The selected features were then used solely for the evaluation of the corresponding test subset. Using this design, access to the test data was prevented during both feature selection and model training, thereby eliminating the risk of leakage.

All the analysis steps were defined and executed using Orange’s visual workflow system. This approach ensured that the analysis could be reproduced exactly, as rerunning all steps with the same data and parameters would yield identical results. The feature-selection method, along with the names of the selected variables and analytical framework, is presented clearly in this study to uphold the principles of transparency and reproducibility.

2.2. Statistical analysis

Sample size calculation: In this study, the lesion diameter was considered the primary characteristic for determining the sample size. According to previous studies,^[13,14] the standard deviation (σ) of the lesion diameter ranged from 9.2 to 15.8 mm. Thus, we assumed σ ≈ 12.5 mm for our calculations. The minimum sample size was calculated with 80% power (α = 0.05, β = 0.20), an effect size (d) of 5 mm, and a Z value of 1.96 for a type I error rate of 0.05. Using the formula n = Z²σ²/ d², the minimum required sample size was determined to be 24.

Statistical analyses were performed using IBM SPSS Statistics version 20.0 (SPSS Inc., Chicago). The age distributions between the NSCLC and round pneumonia groups were compared using the Student t test, while categorical distributions (e.g., sex) were compared using the chi-square test. The performance of the machine-learning classification models was evaluated by calculating the area under the ROC curve (AUC), classification accuracy (CA), sensitivity, and specificity as part of a receiver operating characteristic (ROC) analysis.

3. Results

A total of 48 participants were included in the study and evenly divided into 2 groups: 24 patients with lung tumors and 24 patients with round pneumonia. In the tumor group, 83.3% were male and 16.7% were female; in the pneumonia group, 66.7% were male and 33.3% were female. The chi-square test showed no significant difference in sex distribution between the 2 groups (P = .097). The mean age was significantly higher in the tumor group (65.23 ± 12.30 years) compared to the pneumonia group (37.70 ± 13.55 years), and this difference was statistically significant (P < .001) (see Table 1). In addition to age differences, lesion size was also compared. Regarding the lesion size observed on axial CT scans, the mean lesion diameter was 36 ± 16.3 mm in the tumor group and 34.5 ± 9.4 mm in the pneumonia group. This difference was not statistically significant (P = .71).

Table 1.

Comparison of demographic characteristics of the groups.

		Mass [n = 24]		Pnömonia [n = 24]		P
		Sayi	%	Sayi	%
Gender	Female	4	16.7	8	33.3	.097*
	male	20	83.3	16	66.7
Age, median (IQR)		65,23 ± 12,30		37,70 ± 13,55		<.001**

Chi-square analysis and Mann–Whitney U test were applied.

Bold italic value indicate statistical significance.

IQR = interquartile range.

*P < .05; **P < .01.

Using all radiomic features obtained from both radiologists’ segmentations, high AUC, accuracy, sensitivity, and specificity values were achieved (Tables 2 and 3). In the machine-learning analysis of the first radiologist’s dataset, the Naïve Bayes, SVM, and Random Forest models achieved an AUC of 1.000. In the analysis of the second radiologist’s dataset, the SVM and Naïve Bayes models achieved AUC values of 1.000 as well. For both radiologist datasets, the highest sensitivity and specificity values were observed using the Decision Tree model. The Decision Tree model yielded an AUC of 0.980, a CA of 0.979, a sensitivity of 0.958, and a specificity of 1.000 (Fig. 3).

Table 2.

Machine-learning model analysis results based on the first radiologist’s (HG) radiomic dataset.

Model	AUC	CA	Sensitivity	Specificity
SVM	1.000	0.917	0.958	0.875
Naïve Bayes	1.000	0.896	0.875	0.917
Random forest	1.000	0.896	0.875	0.917
Tree	0.980	0.979	0.958	1.000
Neural network	0.960	0.938	0.917	0.958
Logistic regression	0.807	0.708	0.708	0.708
k-NN	0.736	0.625	0.542	0.708

AUC = area under the curve, CA = classification accuracy, k-NN = k-nearest neighbors, SVM = support vector machine.

Table 3.

Machine-learning model analysis results based on the second radiologist’s (MY) radiomic dataset.

Model	AUC	CA	Sensitivity	Specificity
SVM	1.000	0.917	0.958	0.875
Naïve Bayes	1.000	0.917	0.917	0.917
Random forest	0.992	0.938	0.917	0.958
Tree	0.980	0.979	0.958	1.000
Neural network	0.960	0.917	0.875	0.958
Logistic regression	0.831	0.729	0.667	0.792
k-NN	0.789	0.729	0.583	0.875

AUC = area under the curve, CA = classification accuracy, k-NN = k-nearest neighbors, SVM = support vector machine.

Figure 3.

Receiver operating characteristic curve of the Decision Tree model using all radiomic features.

Using radiomics data, the 5 most important features were identified via information-gain: inverse variance, maximum probability, median, normalized dependence, and long-run emphasis. Using the radiomic dataset derived from these top 5 features, the highest sensitivity and specificity values were achieved with Naïve Bayes, SVM, and Decision Tree models (Table 4). In particular, the Naïve Bayes model achieved an AUC of 1.000, a CA of 0.979, a sensitivity of 0.958, and a specificity of 1.000 (Fig. 4).

Table 4.

Machine-learning model performance using the top 5 radiomic features from the first radiologist’s dataset.

Model	AUC	CA	Sensitivity	Specificity
Naïve Bayes	1.000	0.979	0.958	1.000
Random forest	1.000	0.958	0.917	1.000
SVM	1.000	0.938	0.917	0.958
Tree	0.980	0.979	0.958	1.000
Neural network	0.960	0.979	0.875	0.958
k-NN	0.935	0.917	0.875	0.958
Logistic regression	0.920	0.583	1.000	0.167

AUC = area under the curve, CA = classification accuracy, k-NN = k-nearest neighbors, SVM = support vector machine.

Figure 4.

Receiver operating characteristic curve of the Naïve Bayes model using the top 5 radiomic features.

4. Discussion

Our findings demonstrate that machine-learning models based on radiomic features extracted from CT images can achieve extremely high accuracy in distinguishing round pneumonia from primary lung cancer. When all features were used, the Naïve Bayes and SVM algorithms achieved perfect classification performance (AUC = 1.000) on the datasets segmented by both radiologists. Moreover, the Decision Tree model exhibited the highest sensitivity (95.8%) and specificity (100%), indicating balanced and superior diagnostic performance.

These results are clinically important because round pneumonia is a rare form of pneumonia in adults (constituting <1% of all cases) and can closely mimic the radiological characteristics of primary lung cancer. This overlap can cause diagnostic uncertainty, particularly in atypical presentations, and may lead to unnecessary invasive procedures. Hu et al developed a CT-based radiomic model to differentiate pulmonary cryptococcal granulomas from lung adenocarcinoma and reported a promising diagnostic performance (AUC = 0.94).^[15] Similarly, Xu et al demonstrated that radiomics could effectively predict malignancy in pulmonary nodules of various sizes, highlighting the versatility of such approaches in clinical decision making.^[16] Additionally, while previous studies such as that by Jiang et al focused on using CT radiomics to distinguish among adenocarcinoma subtypes, our study extends this approach to differentiate an inflammatory lesion (round pneumonia) from malignant lesions.^[17] Even advanced imaging modalities, such as F-18 FDG positron emission tomography/CT, have been reported to be unable to reliably distinguish round pneumonia from malignant lesions.^[18]

When the top 5 radiomic features were selected using information-gain, a similarly high performance was achieved. The Naïve Bayes model trained on these 5 features attained an AUC of 1.000, CA of 97.9%, sensitivity of 95.8%, and specificity of 100%, thus underscoring the strong discriminative capacity of the selected parameters. Most of these features reflect gray-level distribution and textural heterogeneity within the lesion. For example, the median pixel value provides an overall estimate of the lesion density. GLCM-based features such as Inverse Variance and Maximum Probability capture the degree of intensity variation and the dominance of certain intensity patterns, while higher-order features, such as normalized dependence and long-run emphasis, describe the homogeneity and continuity of low-intensity regions within the lesion These quantitative parameters accentuate the structural differences between malignant lesions and round pneumonia.

Compared with the greater heterogeneity typically seen in lung cancer, the more homogeneous internal structure and lower gray-level variability of round pneumonia may explain the significant differences observed in these key radiomic parameters. By leveraging these radiomic signatures, our models could distinguish benign from malignant lesions with nearly perfect accuracy. The high-performance metrics (AUC, accuracy, and sensitivity) achieved in this study are consistent with findings from similar radiomics-based research. This study emphasizes the potential of AI-assisted decision support systems in distinguishing rare lesions, such as round pneumonia, which is often misinterpreted as malignancy. However, multicenter studies with larger patient cohorts are required to improve the generalizability of these findings.

Supporting literature also shows that radiomic analysis has been successfully applied to differentiate infectious from malignant pulmonary lesions. Feng et al developed a CT-radiomic nomogram to distinguish solitary pulmonary tuberculomas from lung adenocarcinomas, achieving an AUC of 0.87 in external validation.^[19] Zhao et al combined clinical, radiologic, and radiomic data to differentiate pulmonary cryptococcosis from adenocarcinoma and reported AUC values of 0.91 and 0.89 in the training and test sets, respectively.^[20] Wei et al used conventional CT findings and radiomic features to distinguish mass-like pulmonary tuberculosis from peripheral lung cancer, obtaining an AUC of 0.98.^[21]

Recent studies that have focused on distinguishing pneumonic infections from malignancies further reinforce our results. Yu et al constructed a nomogram incorporating clinical and CT-radiomic variables to differentiate pneumonia-type invasive mucinous adenocarcinoma from typical pneumonia, achieving an external validation AUC of 0.85.^[22] Their study demonstrated that radiomic models can significantly improve the diagnostic accuracy, especially in uncertain cases. Similarly, Selvam et al applied radiomics using a decision tree and XGBoost classifiers to differentiate between benign and malignant small pulmonary nodules, achieving accuracy rates of up to 89%. They found that texture features were the most effective parameters for classification, which is consistent with our findings that underscored the importance of texture-based features for accurate classification.^[23]

In summary, the high-performance metrics achieved in our study (AUC, accuracy, and sensitivity) are in line with similar radiomics-based investigations in the literature. The findings suggest that AI-supported decision systems could have clinical value in accurately classifying lesions, such as round pneumonia, which is rare in adult patients and often mistaken for malignancy. This approach can enhance the diagnostic accuracy in ambiguous or atypical cases and potentially prevent unnecessary invasive procedures by increasing the diagnostic confidence of clinicians.

However, given limitations such as the study’s limited sample size and single-center design, caution is warranted when generalizing the results to broader clinical settings. Although the modeling process employed nested cross-validation to minimize the risk of data leakage, the relatively small number of cases meant that the generalizability of the findings to general clinical practice should be interpreted carefully. Future studies with larger patient cohorts, multicenter designs, and prospective methodologies are needed to validate and extend our results.

5. Conclusion

Our study demonstrated that integrating CT-derived radiomic features with machine-learning algorithms can provide a highly accurate differential diagnosis between round pneumonia and primary lung cancer. This result adds valuable evidence to support the use of radiomic data for noninvasive differentiation of infectious versus malignant pulmonary conditions. The potential clinical applications of this model are promising. In radiology departments, particularly for imaging studies involving numerous nodules, AI-assisted tools can significantly reduce the workload of radiologists, allowing them to focus more effectively on critical or ambiguous cases. The integration of such a decision support system into an existing PACS or imaging software is feasible. For example, when a chest CT scan is uploaded, the system can automatically analyze the detected lesions and provide a risk score to estimate the likelihood of finding round pneumonia versus primary lung cancer. This additional quantitative assessment layer can enhance radiologists’ diagnostic confidence, especially in scenarios in which traditional findings are inconclusive. Moreover, this kind of hybrid decision-making approach can serve as a training aid for less-experienced radiologists by reinforcing structured diagnostic reasoning and improving reporting quality. Previous research has shown that radiomics-based decision support tools can assist in classifying small pulmonary nodules and contribute to earlier and more accurate diagnosis.

6. Limitations

This study had several limitations. First, it was conducted retrospectively at a single center, which may limit the generalizability of our findings to other clinical settings. Additionally, the relatively small sample size of only 48 patients limits the statistical power and robustness of machine-learning models. The limited number of cases also reduced the ability to capture the full spectrum of imaging variability observed in clinical practice. Therefore, future studies with larger patient cohorts, multicenter designs, and prospective methodologies are required to validate and expand upon our findings.

Author contributions

Validation: Hasan Genç.

Visualization: Hasan Genç, Mustafa Yildirim.

Writing – original draft: Hasan Genç.

Writing – review & editing: Hasan Genç.