Abstract
3235
Introduction: There are different parathyroid disorder including parathyroid adenoma, hyperplasia, and carcinoma. Parathyroid adenoma is main cause (about 85%) of hyperparathyroidism. Parathyroid scintigraphy is an efficient approach to distinguish adenoma: and in fact SPECT/CT is advantageous over planar imaging, enhancing diagnostic confidence. We aimed to enable automated discrimination of adenoma from normal cases by using radiomics features and machine learning techniques from delayed parathyroid scans.
Methods: Ninety-two patients (58 adenoma, 34 normal) were studies in this work. All patients underwent parathyroid scintigraphy with injection of 740 MBq of 99mTc-sestamibi and 2 hour delayed SPECT/CT images were acquired. The thyroid region on SPECT/CT images was segmented (with anatomical guide of CT) by an experienced nuclear medicine physician, and radiomic features were extracted in LIFEx v7.0 with IBSI. Images were resampled to 3.296 mm3 isotropic voxel size, and intensity gray level were discretized into 64 fixed bins. Out of 65 radiomics features, 33 were first order and 32 of them were texture features. At first step, patients were split into training (70%) and test (30%) datasets. To remove irrelevant features, Maximum Relevance Minimum Redundancy (MRMR), Recursive Feature Elimination (RFE) and Boruta algorithms feature selection were implemented. Six different machine learning (ML) models including Logistic Regression (LR), eXtreme Gradient Boosting (XGB), Gradient Boosting (GB), Random Forest (RF), Support Vectors Machines (SVM) and K-Nearest Neighbors (KNN) were implemented. Hyper-parameter optimization of ML was performed on the training dataset by 3-fold cross-validation, with optimal models derived on the training data. Performances of models were evaluated on test data, including 1000 bootstraps, to report mean±SD of area-under-the-curve (AUC), accuracy (ACC), sensitivity (SEN) and specificity (SPE) were used to assess the performance of models.
Results: Based on our results (see figure), RFE+XGB had the highest portion of AUC (AUC: 0.76±0.08, ACC: 0.64±0.078, SEN: 0.72±0.1 and SPE: 0.52±0.15). MRMR+GB had highest ACC (AUC: 0.6±0.077, ACC: 0.72±0.072, SEN: 0.85±0.081, SPE: 0.5±0.13) and RFE+SVM had highest SEN (AUC: 0.67±0.094, ACC: 0.71±0.08, SEN: 0.94±0.055 and SPE: 0.31±0.14). Boruta+SVM was top model in SPE (AUC: 0.74±0.092, ACC: 0.65±0.08, SEN: 0.54±0.11 and SPE: 0.82±0.12). Overall, the KNN classifier linked with RFE feature selection had balanced performance in metrics (AUC: 0.63±0.085, ACC: 0.63±0.069, SEN: 0.62±0.1 and SPE: 0.65±0.13).
Conclusions: Radiomic features extracted from delayed parathyroid SPECT combined via machine learning were potentially to differentiate adenoma and healthy cases and achieved good results. SPECT radiomics based model could be potentially implemented in clinics for adenoma detection.