Automatic Interpretation of 18F-Fluorocholine PET/CT Findings in Patients with Primary Hyperparathyroidism: A Novel Dataset with Benchmarks

Primary hyperparathyroidism (pHPT) is a prevalent endocrine disorder characterized by excessive production of parathyroid hormone due to hyperactive parathyroid glands. This paper aims to enhance the diagnostic accuracy of PET/CT using F-fluorocholine (FCH) by employing radiomic image analysis to differentiate hyperfunctioning parathyroid glands (HPTG) from thyroid gland (TG) normal or adenomatous tissue. To this aim, the paper contributes a novel, real-life dataset (made publicly available online for download) along with several benchmarks. First, we collected and labeled FCH PET/CT images from 92 patients with pHPT, extracting 56 s-order and higher-order radiomic features using the software LIFEx. These features were analyzed in 98 HPTG and 91 TG findings, comparing clinical characteristics via non-parametric Wilcoxon rank sum tests. Then, several variants of pattern recognition models (k-nearest neighbor and random forest) and deep neural networks were applied to the task of discriminating between HPTG and TG over the dataset in order to fix baseline results for the Community to challenge. Moreover, two ensemble methods are proposed that combine the aforementioned classifiers, achieving an area under the curve of up to 92.30%. In conclusion, the present integration of radiomic features and machine learning provides a promising approach to the task, setting a benchmark for future research in the field.