Endocrine Abstracts

Non-response and early relapse after peptide receptor radionuclide therapy (PRRT) remain clinically significant in neuroendocrine tumours (NETs). Conventional PET metrics (e.g., SUV_max) incompletely capture tumour biology, whereas pre-treatment somatostatin receptor (SSTR) PET-CT radiomics and machine-learning (ML) may detect subtle phenotypes associated with poor outcomes. Early studies show that radiomic signatures and heterogeneity measures outperform single-voxel metrics for prognostication before PRRT (Laudicella et al., 2022; Wernet et al., 2019; Lee et al., 2023). We conducted a structured review to evaluate whether pre-treatment SSTR PET-CT radiomics/ML features (beyond SUV_max) predict non-response or early relapse after PRRT. Searches were performed in PubMed, Embase and Web of Science for articles published January/2005-September/2025 using combinations of: neuroendocrine, PRRT, SSTR PET, DOTATATE/DOTATOC, radiomics, machine learning, response, progression. After duplicate removal, 196 records were identified; 102 titles/abstracts were screened; 38 full texts were assessed; and 22 studies were included in the qualitative synthesis. Across included studies, radiomics consistently captured intra- and inter-lesional heterogeneity that correlated with progression-free survival (PFS) or objective response, frequently outperforming SUV_max alone. Notably, pre-treatment texture/shape features and lesion-aggregation strategies improved prediction of progression and non-response; heterogeneity indices predicted outcome independent of absolute uptake (Werner et al., 2019). Several investigations reported that global tumour burden and low-uptake voxels (e.g., SUVmin) or volumetric metrics yielded stronger prognostic value than SUVmax, aligning with the biological premise that resistant subclones drive early failure (Lee et al., 2023). Development cohorts using 68Ga-DOTATATE/-TOC radiomics achieved promising discrimination for PRRT response, supporting feasibility of pre-treatment ML triage models (Laudicella et al., 2022). Limitations include small, single-centre datasets; heterogeneous acquisition/reconstruction; variable feature definitions; and limited external validation—factors that can inflate performance estimates and hinder clinical transferability. Implications: AI models trained on harmonised, multi-centre SSTR PET/CT with standardised radiomics pipelines could flag patients at high risk of non-benefit, improving selection and limiting unnecessary toxicity. Next steps should prioritise pre-registered analysis plans, external validation, integration with clinical/biomarker data, and decision-curve analyses to quantify net clinical utility.