Lung cancer conspicuously dominates as the principal male malignancy in terms of incidence and mortality rates, and its trajectory in female cancer incidence is on an unmistakable upward trend (1). Globally, approximately 2.2 million lung cancer cases emerged in 2020, culminating in roughly 1.8 million fatalities (1, 2). Thus, the early detection of malignant lung tumors and timely clinical intervention can effectively reduce mortality and improve patient prognosis. Low-dose computed tomography (CT) has garnered recognition as a potent tool for early lung cancer detection and screening, presenting a statistically significant reduction in lung cancer mortality of approximately 20% when juxtaposed with conventional chest X-ray examinations (3). Nonetheless, the propensity for false positives in CT-based lung cancer diagnosis remains significant (4), with histopathological examination retaining its gold standard status for the diagnosis of malignant lung lesions.

Considering that ultrasound can provide characteristic insights into lesions near the pleura, and the emergence and maturity of contrast-enhanced ultrasound (CEUS) technology, researchers have recently gone beyond simple grayscale ultrasound to deeply study the microvascular supply distribution and dynamic perfusion process of peripheral pulmonary lesions (PPLs) to clarify their vascular characteristics (5–7). Although some studies have used quantitative and qualitative parameters to effectively diagnose benign and malignant PPLs (8–10), few have comprehensively analyzed and evaluated the weight of multiple features of PPLs on grayscale ultrasound and CEUS.

Random forest is a machine learning algorithm that is an ensemble learning model composed of multiple decision trees. Compared with logistic regression analysis, random forest has significant advantages in dealing with feature correlation, feature selection, and handling nonlinear data (11). Random forest models are more appropriate for datasets with many features.

In this study, the clinical data and features of patients with peripheral lung lesions on grayscale ultrasound and CEUS were prospectively collected. A model was established and validated using the random forest algorithm to evaluate the correlation between the relevant variables and the diagnosis of benign and malignant PPLs, thereby revealing the significant factors that distinguish them.

In comparison to previous research methods, we have collected a relatively large amount of general clinical data and ultrasound manifestations of 22 features of PPLs, and extracted 17 features with discriminatory significance between benign and malignant PPLs using logistic regression (9). This increases the likelihood of selecting the optimal features. In comparison to establishing a common multivariate logistic regression model, the advantages of random forests are less susceptible to the influence of multicollinearity between features and can reduce the model’s dependence on a specific training set. Most importantly, random forests are highly useful for situations with a large number of features, helping us to identify the most important features (11). Based on these features, we can comprehensively judge the nature of the lesion, providing a reference for differential diagnosis. We constructed a random forest model in this study based on these 17 indicators to predict the nature of PPLs, with an AUC of 0.92 and an accuracy rate of 88.2% in the test set.

Owing to the dual blood supply characteristics of the lungs, the contrast agent sequentially passes through the right heart, pulmonary circulation, left heart, and systemic circulation after injection into the elbow vein. Consequently, pulmonary artery enhancement occurs earlier than bronchial artery enhancement. Therefore, judging the blood supply source of the lesion by the beginning of the enhancement time can differentiate benign PPLs from malignancies (12). The 2017 ultrasound contrast imaging guidelines released by the European Society of Ultrasound in Medicine and Biology suggest that an initial enhancement time of less than 10 s indicates that the lesion is mainly supplied by the pulmonary artery, which is more common in patients with pneumonia. An enhancement time greater than 7.5 s indicates the possible presence of lung cancer (13). Although benign and malignant PPLs can exhibit different initial enhancement times, an overlap in time between the two may exist, making it impossible to determine the benign or malignant nature of PPLs. In our study, when the lesion AT was greater than 6.05 s, the likelihood of malignant lesions increased. However, due to the different physiological or pathological conditions of patients (such as examination location, chronic heart failure, hyperthyroidism or hypothyroidism, and chronic lung diseases), inconsistent injection speeds of the contrast agent by operators, or even uncoordinated timing by operators, the beginning enhancement times of pulmonary artery and bronchial artery enhancement in patients can be variable (14). Therefore, relying solely on lesion AT for judgment carries risks; thus, the importance ranking in the random forest variables is not high. Therefore, we introduced the difference in the time of arrival of the lesion. When the lesion-lung AT difference was greater than 2.45 s, malignant lesions can be diagnosed. However, in some patients, the difference between the lesion and lung AT values was lower than the critical value. In contrast, the difference between the lesion and chest wall AT values was very close. In our study, such lesions were classified as benign; however, their actual blood supply was closer to that of the chest wall. In this study, we found that the variable importance ranking in the random forest was the first AT difference ratio, which is consistent with the findings of Bi et al. (10, 15). The AT difference ratio was more effective in reducing individual differences than the lesion-lung AT difference. When the AT difference ratio was > 0.45, the likelihood of malignant tumors increased. In addition, statistically significant differences were observed in TTP between the benign and malignant groups, whereas no statistically significant differences existed in RT between the malignant and benign groups, which is consistent with the studies by Bi et al. (15) and Li et al. (9). However, the TTP may also be affected by the lesion AT.

This study was limited to distinguishing between benign and malignant lesions only through quantitative indicators, and we also need to analyze the enhancement patterns of the lesions. In benign lesions, the vascular structure is usually tree-like, the perfusion is unidirectional and radiating, and it shows uniform and significant enhancement. In malignant lesions, the vascular appearance is mostly dotted and curly, and it shows uneven and low enhancement, which is similar to the findings of Li et al. (9) and Shen et al. (16). In benign lesions, the supplying pulmonary artery usually maintains its normal structure; therefore, it presents a tree-like structure radiating from the hilum to the pleural direction on CEUS. Due to inflammatory reactions, vascular dilation, and acceleration of blood flow, the local blood flow increases, resulting in the lesion mostly showing high enhancement. Malignant tumor cells destroy the normal structure of the lung tissue, and the small bronchial arteries gradually replace the hypoxic pulmonary arteries as the tumor grows. The structure of the disordered new vessels produces a large number of anastomotic branches and collaterals, and the distribution is tortuous and uneven, resulting in the enhanced vessels in the lesion showing a dotted or curly appearance and low enhancement (17). However, some studies have shown that the degree of enhancement of malignant lesions is related to pathological results. According to the findings of Findeisen et al. (18) and Wang et al. (19), compared with squamous cell carcinoma, adenocarcinoma has a higher expression of vascular endothelial growth factor, resulting in a higher microvascular density in adenocarcinoma and a higher degree of enhancement. In our study, adenocarcinoma accounted for a higher proportion of tumors in the malignant group, which may have led to malignant lesions showing isoenhancement rather than low enhancement. In addition, we found that the convergent perfusion pattern was more common in malignant lesions, which is consistent with the research of Bi et al. (15) and Fu et al. (20), whereas a local-to-global perfusion pattern was more common in the study by Shen et al. (16). However, these perfusion patterns are related to tumor damage to the blood vessels and the formation of new blood vessels. Moreover, the number and morphology of necrotic areas were significantly different between the benign and malignant groups. However, these qualitative imaging indicators were not highly ranked in the random forest models used in the present study.

The size and shape of the lesion ranked second and fourth, respectively, in terms of the variable importance of the random forests. Unlike previous studies, our clinical experience led to a more detailed description of the shape, which is similar to the results reported by Li et al. In the present study, wedge-shaped lesions were more common among benign lesions, whereas round, spindle, and irregular shapes were more common among malignant lesions. Pneumonia can cause inflammatory cell infiltration and tissue necrosis in local tissues, resulting in vascular occlusion and ischemia of lung tissue in the affected area, thus forming wedge-shaped lesions with blurred boundaries. Malignant cells grow rapidly and spread through blood vessels and bronchi, resulting in increased longitudinal diameter. During growth, surrounding tissues can exert a certain tension, resulting in larger malignant tumors in a ball, spindle, or irregular shape, with clear lesion boundaries (20, 21). The research by Gould et al. (22) shows that the size of the lesion is a key factor, and that the larger the diameter of the lesion, the higher the possibility of malignant lesions. However, this is inconsistent with the studies of Bai et al. (23) and Bi et al. (10), which may be related to small sample datasets or different indicators used to describe the size of the lesion. In this study, we used the sum of the horizontal and vertical diameters to reflect lesion size, considering the size of the lesion in two directions, which can more accurately assess the overall size of the lesion compared with only describing the size of the lesion using the horizontal, vertical, or maximum diameters.

In our study, age ranked third in terms of variable importance in the random forest model, which is in line with the increasing trend in cancer incidence with age. At the same time, smoking is an important risk factor for cancer (1), but no significant correlation was found between smoking and benign and malignant tumors in this study. This may be due to the small sample size or the different proportions of pathological results in the collected cases. According to the study by Loeb et al. (24), smoking is closely related to lung cancer (especially squamous cell carcinoma), but not adenocarcinoma. Importantly, adenocarcinoma accounted for approximately half of the malignant tumors in the present study, which could explain the lack of a significant difference between smoking and benign and malignant tumors.

The limitations of this study are as follows: (1) the number of cases included in this study was small, and the ability of grayscale ultrasound and CEUS to differentiate benign and malignant PPLs still needs further validation with multicenter trials with large sample sizes; (2) the placement of the CEUS region of interest is subjective; (3) this study used ultrasound-guided biopsy, which may lead to sampling errors or insufficiencies in the lesion; and (4) this study did not further analyze the lesion subtypes.

The random forest algorithm model, based on clinical data and ultrasound imaging features, has clinical application value in predicting benign and malignant PPLs. Combining indicators such as AT difference ratio, size, age, and shape of the lesion can provide a diagnostic basis for differentiating benign and malignant PPLs.

The raw data supporting the conclusions of this article will be made available by the authors, without undue reservation.

The studies involving humans were approved by Medical Ethics Committee of the Second Affiliated Hospital of Harbin Medical University. The studies were conducted in accordance with the local legislation and institutional requirements. Written informed consent for participation in this study was provided by the participants’ legal guardians/next of kin. Written informed consent was obtained from the individual(s), and minor(s)’ legal guardian/next of kin, for the publication of any potentially identifiable images or data included in this article.

HW: Data curation, Methodology, Writing – original draft. YCW: Data curation, Writing – original draft, Conceptualization, Writing – review & editing. JL: Data curation, Writing – review & editing, Formal analysis, Investigation. YYW: Data curation, Investigation, Writing – review & editing. LL: Data curation, Formal analysis, Writing – review & editing. JS: Writing – review & editing. XW: Writing – review & editing.