Investigation of the efficacy of radiomics parameters obtained from computed tomography imaging in detecting the primary tumor origin of metastatic lung lesions

Objective: Metastatic nodules can be detected in patients undergoing chest computed tomography (CT) for reasons other than malignancy surveillance. These cases are classified as malignancies of undefined primary origin. Metastatic nodules often exhibit similar radiological appearances, making it challenging to identify the primary tumor site using conventional imaging characteristics such as shape, margins, density, and contrast enhancement. Consequently, the detection of a malignancy of undefined primary origin often requires a range of diagnostic evaluations, including laboratory tests, advanced radiological imaging, biopsies, histopathological assessments, endoscopic procedures. These examinations are costly and expose patients to the risks associated with invasive procedures, in addition to delaying the initiation of appropriate treatment. Furthermore, in a subset of patients, the primary tumor site remains unidentified despite comprehensive evaluations. Therefore, achieving an accurate differential diagnosis in a shorter time and at a lower cost is of paramount importance. Radiomics, a method that extracts quantitative data from medical images via mathematical algorithms, offers a potential solution. In this study, we aimed to differentiate lung metastases from breast cancer, colorectal cancer, and renal cell carcinoma—among the most common metastatic tumors to the lung—using radiomic features and various clinical-radiological parameters. Materials and Methods: A total of 70 metastatic breast cancer, 85 metastatic colorectal cancer, and 74 metastatic renal cell carcinoma patients diagnosed at two hospitals within the same medical center between January 1, 2015, and December 31, 2023, were included in the study. One hospital served as the training cohort (160 patients), while the other was designated as the test cohort (69 patients). The initial chest CT scans, in which metastases were detected, were analyzed, and clinical-radiological parameters—including patient age and sex, number of metastases, laterality, presence of mediastinal lymphadenopathy, bone and liver metastases, lymphangitic spread, pleural effusion, and pleural metastases—were recorded. Five variables (age, laterality, number of metastases, bone metastasis, and liver metastasis) showed statistically significant differences among the three groups (p<0.05). For radiomics analysis, non-cavitating nodules larger than 5 mm in contrast-enhanced CT scans were segmented three dimensionally. Following preprocessing, 854 radiomic features were extracted, including 110 original and 744 filtered features. Features with an intraclass correlation coefficient (ICC)<0.9 were excluded, leaving 441 features for further analysis. To classify the three groups, the Least Absolute Shrinkage and Selection Operator (LASSO) regression was applied, yielding eight features with nonzero coefficients. Random Forest (RF), Support Vector Machine (SVM), and XGBoost algorithms were used for modeling, with performance assessed using the area under the curve (AUC). Results: Three different models were developed: clinical-radiological models based on the five statistically significant variables, radiomics models incorporating the eight features identified through LASSO regression, and combined models using both clinical-radiological variables and radiomics features. Since RF was the most successful algorithm for radiomics-based classification, it was chosen for comparative analysis. The macro-/micro-averaged AUC values for the RF models were 0.616/0.616 for the radiomics model, 0.694/0.623 for the clinical-radiological model, and 0.714/0.712 for the combined model. Conclusion: Radiomics features alone demonstrated limited effectiveness in distinguishing lung metastases from breast cancer, colorectal cancer, and renal cell carcinoma. This situation may stem from the structural similarities among metastatic nodules on contrast-enhanced imaging and the insufficiency of harmonization processes. This hypothesis can be tested with larger patient cohorts. The most successful model in distinguishing metastatic nodules was the combined model, which showed moderate-to-high performance. Keywords: Lung metastasis, malignancy of undefined primary origin, computed tomography, radiomics