A radiomics approach to discriminate lung cancer from pneumonia on FDG PET-CT

Objectives: ackground Infectious pneumonia and primary lung cancer often show similar image findings. The clinical course and FDG PET/CT help in differentiation, but sometimes it is not enough for even experience nuclear medicine physician. Overdiagnosis may cause unnecessary invasive tests or surgery, and thus strict differentiation is required. Recently, texture analysis on PET imaging is applied for oncology studies. Texture features reflect the heterogeneity of tumor metabolism which could be helpful for differential diagnosis from inflammation disease. In this study, we investigated the performance of single and multiple parameters of texture analyses using support vector machine (SVM) for discriminating primary lung cancer from pneumonia from PET/CT images. Methods ¹⁸F-FDG PET/CT images of the latest consecutive 20 patients of non-small cell lung cancer (LC) and pneumonia (PN), respectively, were retrieved from our hospital database. All the LC patients were pathologically proven. We confirmed all the PN lesions had disappeared in the clinical course. Scanner was either a Siemens Biograph 64 PET-CT scanner or a Philips GEMINI TF-64 scanner. For each lesion, 3 delineation methods were applied: Nestle method (adaptive threshold method by Nestle et al, β=0.3), SUV2 method (VOIs with a fixed threshold of SUV≥2.0), and Liver3SD method (VOIs with threshold of SUV of the liver mean + 3 SD). In addition, background texture was computed from VOI in the normal lung, the liver and the muscle. The voxel intensities were resampled using 64 discrete values, between minimum and maximum SUVs (min-max) or between SUV 0 to 20 (SUV0-20). In addition to histogram analysis, 4 texture matrices (gray-level co-occurrence matrix (13 directions), gray-level run length matrix (13 directions), gray-level zone size matrix, and neighborhood gray-level difference matrix) were generated to calculate a total of 36 texture parameters. SUVmax, SUVmean, MTV, TLG were also measured. SVM with linear kernel was employed for machine learning. Single or multiple texture features were given to SVM to classify the image to lung cancer or pneumonia. Accuracy was estimated using leave-one-out validation. Image analysis was performed on Metavol. Texture features were computed using ‘ptexture’ on Python. Results When a single texture parameter was given to SVM, the accuracy was 40% to 75% (dissimilarity and LGRE reached 75%; SDhist, homogeneityGLCM, contrastGLCM, and LZHGE reached 70%). When the full set of 40 parameters of the lesion was given to SVM, the accuracy was degraded to 60% (min-max) to 65% (SUV0-20). When a lesion-and-reference combination (i.e., 80 parameters) was given to SVM, the accuracy was maximized to 82.5%, which was achieved by Liver3SD method and the liver texture. Finally, the entire texture matrix (240 parameters per patient) was given to SVM, and the accuracy was 40% to 67.5%. Conclusion Support vector machine with multiple texture features on FDG PET-CT discriminated lung cancer from infectious pneumonia with accuracy of 82.5%. The machine learning system may provide additional information to physicians in interpreting images, although parameter selection needs to be optimized.