FDG-PET (fluorodeoxyglucose positron emission tomography) is an imaging technique that measures metabolic activity in tissues. Cancer cells, which generally consume more glucose than normal cells, take up more FDG and appear as bright spots on PET scans. In pancreatic cancer, FDG-PET can reveal the metabolic extent of disease that may not be visible on CT alone.
Standard metrics extracted from PET scans -- such as the maximum standardized uptake value (SUVmax) or total lesion glycolysis (TLG) -- provide some prognostic information but may miss the spatial complexity of tumor metabolism. Just as CT texture analysis captures heterogeneity invisible to visual inspection, PET radiomics can extract dozens of features describing how metabolic activity is distributed throughout the tumor volume.
This study investigated whether machine learning applied to a comprehensive set of PET radiomic features could identify which features were most prognostic for overall survival in pancreatic cancer, potentially revealing new imaging biomarkers beyond the simple metrics currently used in clinical practice.
The study enrolled 161 pancreatic cancer patients who underwent FDG-PET imaging. Tumor volumes were manually delineated on PET scans, and 42 radiomic features were extracted from each, covering intensity statistics (like SUVmax, mean, and peak), volumetric measures (like TLG and metabolic tumor volume), and texture features describing spatial heterogeneity of the metabolic signal.
A random forest classifier was trained to identify which features were most informative for predicting overall survival. Random forests rank feature importance using the Gini index, which measures how much each feature reduces impurity when used to split the data at a decision node. Features with high Gini importance contribute more consistently to accurate predictions.
Multivariate Cox proportional hazards regression was used to confirm that the top radiomic features remained independently prognostic after accounting for established clinical variables, including surgical resection status, which is the most powerful determinant of survival in pancreatic cancer.
Patients were stratified into risk groups by combining the most important radiomic feature with TLG, a volumetric PET parameter, to assess whether a combined imaging score could meaningfully separate patients with distinct prognosis profiles.
The random forest analysis identified GLZLM GLNU (Gray Level Zone Length Matrix Gray Level Non-Uniformity) as the most important PET radiomic feature for predicting overall survival. In multivariate Cox regression, GLNU was significantly associated with survival with a hazard ratio of 2.1 (p=0.011), meaning patients with high GLNU had more than twice the risk of death compared to those with low GLNU, even after controlling for other factors.
GLNU measures how unevenly distributed metabolic intensity is across zones of the tumor. A high GLNU indicates that some parts of the tumor are highly metabolically active while others are not -- a pattern of metabolic heterogeneity that reflects biological aggressiveness, poor tumor differentiation, or regions of rapidly proliferating versus necrotic cells.
When patients were stratified using a combined score of GLZLM GLNU plus TLG, three distinct risk groups emerged with significantly different median survival times. This three-tier stratification provided better separation than either feature alone, demonstrating complementarity between the heterogeneity measure and the volumetric metabolic burden.
While radiomic features provided independent prognostic value, surgical resection remained the strongest clinical predictor of survival in this cohort, as expected. Patients who underwent surgery had dramatically better outcomes than those who did not. This clinical variable dominated the multivariate models, and radiomic features were most informative within the context of patients receiving specific treatments.
Within the group of surgically resected patients, the prognostic value of radiomic features was more pronounced, suggesting these imaging biomarkers may be most clinically useful for fine-grained prognostic stratification among patients who are surgical candidates -- where the question is not whether to operate but how intensively to treat after surgery.
The finding that radiomic features add prognostic information independent of surgical status suggests they are capturing tumor biology that influences outcomes regardless of treatment approach, potentially reflecting fundamental properties of tumor aggressiveness that determine metastatic potential and treatment resistance.
The central insight from this study is that the spatial distribution of metabolic activity within a tumor provides more prognostic information than simply measuring the total or maximum metabolic activity. A tumor that is uniformly active metabolically may behave differently from one with irregular hotspots of high and low activity.
Metabolic heterogeneity on FDG-PET likely reflects underlying cellular heterogeneity -- variations in cell density, differentiation, hypoxia, and necrosis across different tumor regions. These biological variations drive differences in how tumors respond to treatment and in their metastatic potential. Capturing this spatial complexity through radiomics extracts clinically meaningful information that was previously inaccessible from routine PET scans.
The combination of GLNU with TLG is practically appealing because both features are extractable from a standard clinical FDG-PET scan without additional imaging or testing. If validated prospectively, this combined metric could be computed routinely from existing scans and incorporated into prognostic models immediately without changing clinical workflow.
This study demonstrates that machine learning applied to FDG-PET radiomic features can identify prognostic biomarkers in pancreatic cancer, with GLZLM GLNU emerging as the most informative single feature and its combination with TLG enabling three-tier patient risk stratification.
The practical value of this approach lies in its accessibility -- FDG-PET is already performed in many pancreatic cancer patients for staging purposes, meaning the same scan can simultaneously provide metabolic staging and a prognostic radiomic signature at no additional patient burden or cost. This is the kind of added-value approach that can be widely implemented once validated.
Future work should prospectively validate these features in independent cohorts, explore their performance in the context of modern combination chemotherapy regimens, and investigate whether baseline PET radiomics can predict response to treatment as well as overall prognosis. If response prediction is possible, these features could guide personalized therapy selection from the initial staging scan.