Evaluation of primary tumor origin in lung metastases with artificial intelligence algorithms

Lungs are one of the most common site of metastasis in malignancies which can be involved in 20-54% of the extrapulmonary tumors. The diagnosis in lung metastases is usually established on computed tomography (CT) with predominantly lower zone bilateral multiple nodular solid lesions. Distinctive imaging features such as cavitation, calcification, hemorrhage, edema and diffuse miliary pattern can be beneficial in differentiating the malignancies that can present with lung metastases. However, in most of the cases, it is usually not possible to differentiate the primary malignancy on CT images. In recent years, image processing methods with artificial intelligence have gained recognition in radiology and other medical specialties. Radiomics has made it possible to capture information from radiological images that may indicate shape, contour or neighboring pixel relationships, which than act like a imaging biomarker that can be used to diagnose, grade, prognosticate the disease outcome and evaluate the response to treatment. In our study, we aimed to evaluate the performance of artificial intelligence models that are trained with radiomics imaging features, to differentiate the primary tumor origin in cases with lung metastases on chest CT images. We searched our hospital database from January 2014 to December 2020 retrospectively, and included 165 cases with 482 metastatic lesions in the study. Cases were arranged in 7 subgroups according to primary tumor origin that comprises colorectal cancers, renal cell carcinomas, bladder cancers, breast cancers, pancreas cancers, prostate cancers and sarcomatous cancers. Imaging studies were performed on Toshiba Aquilion 64-detector (Toshiba Medical Systems, Otawara, Japan) and Siemens Somatom Definition Flash 128-detector (Siemens Healthcare, Erlangen, Germany) CT scanners after intravenous contrast administration. Tumoral lesions were labeled with 3D Slicer (http://www.slicer.org, version 4.11) using the "Segmentation Wizard" semi-automatic segmentation module. Gray scale normalization was applied to tumoral regions as a preprocessing method. Thirty-four first-order and Haar-like features extracted from the images by using free and open-source Python software libraries. To test the model performance, 30% of the data were randomly splitted. K-nearest neighbor (KNN), support vector machine (SVM), random forest (RF) and gradient boosting (GB) models were trained, tested and best performing models were established. Accuracy, precision, recall and F1 score were used to evaluate the model performance. Best accuracy values on the test data were achieved with pancreas and prostate cancer subgroups on KNN model (0.93 and 0.92 respectively), prostate, pancreas and sarcomatous cancer subgroups on SVM model (0.93, 0.90 and 0.90 respectively), pancreas and sarcomatous cancer subgroups on GB model (0.94 and 0.94 respectively), prostate and pancreas cancer subgroups on RF model (0.93 and 0.90 respectively). Best classifier performance with regard to F1 score was achieved in the breast cancer subgroup on KNN, SVM and RF models with 0.62, 0.66 and 0.50 respectively. GB model demonstrated the highest F1 score with 0.77 on the sarcomatous cancer subgroup. GB model F1 score was 0.74 on the breast cancer subgroup. In total, the best F1 scores among the 4 models were achieved with GB and SVM models on the breast and sarcomatous cancer subgroups. Radiomics imaging features and artificial intelligence models can have a potential in differentiating the primary subgroups of lung metastases as a non-invasive diagnostic tool. To achieve that, more studies on large and independent datasets with standardized methods are required. Keywords: Artificial intelligence, CT, lung metastasis, radiomics.