In this study, TARE local response prediction capabilities of ML models based on pre-procedure 18F-FDG PET radiomics and clinical features were assessed in patients with CRCLM. The response groups were categorized into non-viable and viable tumor tissue. Three distinct classifiers (Random Forest, Extreme Gradient Boosting, Logistic Regression) with radiomics-only features demonstrated AUCs ranging from 0.90 (95% CI 0.71-1) to 0.81 (95% CI 0.51-1) in an un-seen test set. Using the clinico-radiomics dataset classifier AUCs varied between 0.88 (95% CI 0.51-1) and 0.84 (0.62-1). In both datasets, RF classifier achieved 100% sensitivity at the optimal threshold. This study represents a promising preliminary analysis of our methodology. A larger, multicenter study confirming the utility of our methodology could significantly impact patient selection in the clinical setting. In other words, before deciding to include patients in the TARE schedule, oncology tumor boards could anticipate the benefit provided by TARE.

Previous studies investigated TARE response prediction predominantly using MR or CT based radiomics in HCC patients [5, 12,13,14,15,16,17,18,19]. Ince et al. [5] reported significant performance difference regarding predicting TARE response in patients with HCC when MR-based radiomic features included to the clinical features. However, Marinelli et al. [12] reported that MR-based radiomic features yielded higher AUC values for response prediction compared to clinical data, in HCC patients undergoing treatment with resin spheres, however, the difference was not significant. The performances of ML models were not significantly altered when the clinical features were combined with the radiomics data, in the present study.

PET-based radiomics were also utilized in the literature for the purpose of TARE response anticipation [20,21,22,23]. When comparing our findings with those of previous PET-based radiomics studies that evaluated TARE response [20,21,22,23], it is important to emphasize several key differences. Unlike most earlier investigations, which focused predominantly on HCC and often incorporated whole-liver radiomics or post-therapy 90Y PET metrics, our study uniquely evaluates pre-procedural 18F-FDG PET radiomics in patients with CRCLM. Previous studies have reported varying predictive performances, with AUC values generally ranging from moderate to high. However, these studies frequently used heterogeneous imaging protocols, multiple scanners or radiomics derived from post-treatment imaging, which may influence feature stability and model performance. In contrast, our use of a single scanner and standardized acquisition protocol probably reduced technical variability, which may explain the high test-set AUCs observed. Furthermore, the distinct tumour biology of CRCLM compared with HCC likely contributes to differences in the identified radiomic predictors. These methodological and biological differences may explain the discrepancy between previous findings and the present results, and highlight the need for multicenter, disease-specific validation. Blanc-Durand et al. [22] reported that pre-treatment whole-liver 18F-FDG PET could predict overall survival and hepatic progression free survival in patients with HCC. Wei et al. [23] published that their model utilizing radiomics features extracted from post-therapy 90Y PET and mean absorbed dose, reached an AUC of 0.803 (95% CI 0.702–0.758) for predicting lesion response in HCC patients. They emphasized that despite the noisy nature of post TARE PET images, combined ML models could predict disease progression.

Radiomics is an emerging and rapidly growing field of research area in radiology [24]. There has been a notable increase in the number of studies reporting outcome predictions through the utilization of radiomics. However, it is imperative to acknowledge certain limitations observed in these studies. The majority of reported studies exhibit variations in machine learning pipelines, the presence of an unseen test set, and imaging protocols. Moreover, the utilization of multiple scanners for image acquisition, heterogenous patient population, and small sample sizes are also commonly seen. The reported high scoring performance metrics might be over optimistic, considering their lack of generalizability and reliability. Nevertheless, radiomics could still provide helpful information in clinical practice. A recent meta-analysis investigating early prediction of radioembolization treatment response with radiomics, have shown a pooled sensitivity and specificity of 83.5% (95% CI 76%-88.9%) and 86.7% (95% CI 78%-92%), respectively [25]. In the current study, when clinical features were included, three of the five radiomic features remained at the top, and radiomic features alone showed similar performance to the combination of radiomic and clinical features in predicting response to TARE, underscoring the reliability of radiomic features. The top five radiomic features identified in the present models may reflect the underlying biological characteristics of the tumor. ‘Coarseness’ quantifies the rate of intensity change across neighboring voxels and is often associated with the degree of intratumoral heterogeneity — coarser textures may correspond to more heterogeneous metabolic activity or necrotic components. IMC1 reflects the complexity and non-uniform relationships among grey-level intensities. Lower IMC1 values have been associated with disorganised microstructural patterns and aggressive tumor behaviour. ZE measures the randomness of homogeneous zones within the tumor and increases with metabolic heterogeneity, which may be associated with uneven cellularity or variable perfusion. SZN describes variability in zone sizes — higher SZN values indicate irregular clusters of similar intensity and may reflect heterogeneous architecture, such as the presence of both viable and necrotic regions. ‘Strength’ captures the perceptual prominence of structured patterns within the lesion. Higher values may be associated with well-defined, spatially coherent metabolic patterns, whereas lower values reflect disorganised tissue structure. While these relationships are not direct histological equivalents, they imply that predictive radiomic features could act as surrogates for tumor heterogeneity, cellular disorganisation, and microstructural complexity.

In a recent CT based study by Roll et al. [26] including patients with CRCLM, found AUC of 0.75 (95% CI 0.48-1). The authors made the radiomics analysis from enhanced CT images obtained from different scanners under different CT parameters. Our study is unique in being 18F-FDG PET radiomics-based TARE response prediction focused solely on patients with CRCLM. Also we utilized a uniform image acqusition, evaluated radiomics data obtained from the same PET scanner with the same imaging protocol for all patients leading to less data heterogenity.

This study has several limitations. First of all, external validation was not performed, and it is unclear whether the results of this study possess generalizability applicable to clinical practice. A relatively small sample size and retrospective nature of the study were the other major drawbacks. The single-center, retrospective design of the study inherently restricts the generalizability of the findings. As all PET/CT examinations were performed on the same scanner using a uniform acquisition and reconstruction protocol, the identified radiomic features may reflect scanner- or protocol-specific characteristics rather than universally reproducible biomarkers. PET/CT radiomics is sensitive to variations in technical parameters and differences in scanners, reconstruction algorithms or segmentation workflows across institutions could affect feature stability and model performance. Therefore, external validation in larger, multicenter cohorts with heterogeneous imaging platforms and standardized acquisition protocols is essential to confirm the robustness and reproducibility of our proposed models. Furthermore, our analysis did not include multiple segmentations with different segmentators, which means that the reproducibility of features was not thoroughly assessed. Lastly, the 3-month follow-up evaluation may have led to misinterpretation such as pseudoprogression due to the early timing of the assessment.

In conclusion, the current study demonstrated the potential of baseline 18F-FDG PET radiomics for predicting TARE response in patients with CRCLM. These findings could inform future studies using larger, standardised datasets, which may lead to the development of a clinical guidance tool for patient selection and personalised medicine.