Positron Emission Tomography Imaging Biomarker and artificial intelligence for the Characterization of Solitary Pulmonary Nodule (SPN)

Ashish Kumar Jha^1,2*Umeshkumar Baburao Sherkhane^1,3Nilendu C. Purandare^1,2Leonard Wee³Andre Dekker³VENKATESH RANGARAJAN^1,2

• ¹Department of Nuclear Medicine and Molecular Imaging, Tata Memorial Hospital, Mumbai, Maharashtra, India
• ²Homi Bhabha National Institute, Mumbai, Maharashtra, India
• ³Department of Radiotherapy, GROW School for Oncology & Reproduction, Faculty of Health, Medicine and Life Sciences, Maastricht University, Maastricht, Netherlands

The characterization of solitary pulmonary nodules (SPNs) as malignant or benign remains a diagnostic challenge using conventional imaging parameters. The literature suggests using combined Positron Emission Tomography (PET) and Computed Tomography (CT) to characterise a SPN. Radiomics and machine learning are other promising technologies which can be utilised to characterise the SPN. Purpose: This study explores the potential of PET radiomics signatures and machine learning algorithms to characterise the SPN. Methods: This retrospective study aimed to characterize solitary pulmonary nodules (SPNs) using PET radiomics. A total of 163 patients who underwent PET/CT imaging were included in this study. A total of 1098 features were extracted from PET images using PyRadiomics. To optimize model performance two strategies i.e. a) feature selection and b) feature reduction techniques were employed, including hierarchical clustering, RFE in feature selection, and PCA in feature reduction. To address outcome class imbalance, the dataset was statistically resampled (SMOTE). A random forest models was developed using original training set (RF-Model-O & RF-PCA-Model-O) and balanced training dataset (RF-Model-B & RF-PCA-Model-B) and validated on the test datasets. Additionally, 5-fold cross-validation and bootstrap validation was also performed. The model's performance was assessed using various metrics, such as accuracy, AUC, precision, recall, and F1-score.Results: Of the 163 patients (aged 36-76 years, mean age 58 ± 7), 117 had malignantmetastatic disease and 46 had granulomatous or benign conditions. In Strategy (a), five radiomic features were identified as optimal using hierarchical clustering and RFE. In Strategy (b), five principal components were deemed optimal using PCA. The model accuracy of RF-Model-O and RF-Model-B in the train-test validation, 5-fold cross-validation and bootstrap validation were found to be 0.8, 0.80 ± 0.07, 0.84 ± 1.11 and 0.8, 0.83 ± 0.10, 0.80 ± 0.07 in Strategy (a).Similarly, the model accuracy of RF-PCA-Model-O and RF-PCA-Model-B in the train-test validation, 5-fold cross-validation and bootstrap validation were found to be 0.84, 0.80 ± 0.07, 0.84 ± 07 and 0.74, 0.80 ± 0.08, 0.75 ± 0.08 in Strategy (b).The PET radiomics demonstrated excellent performance in characterizing SPNs as benign or malignant.