Machine learning based radiomics analysis of SPECT images in predicting neuroendocrine tumors

• Applying radiomics to SPECT imaging to predict Neuroendocrine tumors (NETs) grade. • Adding Machine leanring algorithm to evaulate our prediction accuracy. • Evaluating the importance of the Ki-67 index in cases both with and without grade discrepancy. • Identifying which feature holds greater validity in predicting NET grade and ascertain if there is a different between Ki-67 index and tumor grade with and without discrepancy. Over the past few decades, the prevalence of neuroendocrine tumors (NETs) has significantly increased. Current pathological assessment methods often rely on invasive procedures. This study aims to develop machine learning (ML)-based radiomics models using SPECT images to predict grade 1 and 2 NETs and assess the Ki-67 index, particularly in cases with grade discrepancy. We developed a retrospective research involving 144 patients with pathologically confirmed NETs collected from five nuclear medicine centers. The dataset included patients with available Ki-67 index values and those exhibiting discrepancies between their Ki-67 index and tumor grade (i.e., low Ki-67 index and intermediate grade). Radiomic features were extracted from SPECT images, and feature selection was performed using the Least Absolute Shrinkage and Selection Operator (LASSO) method. Radiomic signature models were then constructed using a decision tree classifier. The prediction of tumor grade in absence of Ki-67 index resulted in area under the curve (AUC) of 0.83, in the presence of Ki-67 index with discrepancy an AUC of 0.86, and in presence of Ki-67 index without discrepancy an AUC of 0.99. The proposed radiomics model based on SPECT images, and machine learning can effectively predict and assess classification of NET grade. Although Ki-67 has been highly recommended as a strong predictive feature in previous studies, our findings suggest that its predictive value diminishes in cases with grade discrepancies, such as low Ki-67 index paired with intermediate grade.