Small-cell lung cancer has one of the worst prognoses of any cancer, with a 5-year survival rate below 7%. The majority of patients already have metastatic disease at diagnosis, and the location and pattern of metastases profoundly shape survival outcomes, with liver, brain, bone, and lung involvement all associated with particularly poor prognosis.
Total tumor burden measured by metabolic tumor volume (MTV) or total lesion glycolysis (TLG) on FDG PET/CT is an established prognostic marker, but it captures only the aggregate amount of disease and not the spatial distribution of metastases across the body, which may carry additional prognostic information that TLG alone cannot convey.
Clinicians qualitatively consider metastasis distribution patterns such as the distance from the primary tumor to the farthest metastatic site or the number of involved organs, but without quantification these observations cannot be consistently integrated into survival models or treatment response assessments.
This study introduces two novel radiomic approaches, TLGd (total lesion glycolysis with whole-body tumor distribution) and METAORG (organ-based metastatic distribution patterns), validated in 520 SCLC patients across internal and external cohorts to determine whether spatial distribution improves prognostic prediction beyond conventional TLG.
520 SCLC patients were enrolled across two hospital cohorts. 364 patients from Shinchon Severance Hospital formed the training (n=291) and internal test (n=73) sets, while 156 patients from Gangnam Severance Hospital formed an independent external test set. All patients underwent baseline FDG PET/CT before treatment.
For TLGd, all tumor lesions including the primary tumor, lymph nodes, and metastases were manually segmented by experienced nuclear medicine physicians using an SUV threshold of 2.5. All lesions were merged into a single combined segment, and radiomic features were extracted from this whole-body tumor mask to capture spatial distribution patterns independent of individual organ boundaries.
For METAORG, the TotalSegmentator deep learning tool automatically delineated eight organs on CT: the liver, spleen, axial skeleton (cervical through lumbar spine and sacrum), peripheral skeleton (proximal extremities, scapulae, and pelvis), core muscles, adrenal glands, thyroid, and lungs. These CT-based organ volumes of interest were then transferred onto PET images for radiomic feature extraction reflecting organ-level FDG distribution.
Two separate machine learning models were developed: a Random Survival Forest for survival duration prediction using the concordance index, and a Random Forest classifier for survival event prediction evaluated by area under the ROC curve. Feature selection combined permutation importance and LASSO regularization from an initial pool of 1,984 features across all data sources.
The combined model incorporating TLG, clinical factors, and radiomics achieved the highest predictive performance in both test sets. In the internal test, the combined model reached a C-index of 0.753, compared to 0.611 for TLG alone and 0.592 for clinical factors alone. In the external test, the combined model C-index was 0.740 versus 0.637 for TLG and 0.326 for clinical factors.
The radiomics-only model (TLGd and METAORG without conventional TLG or clinical data) achieved C-indices of 0.721 (internal) and 0.706 (external), substantially outperforming both TLG-only and clinical-only models. This demonstrates that distribution pattern information is more prognostically informative than either aggregate tumor burden or standard staging information.
Notably, adding clinical data to the radiomics model only marginally improved internal test performance from 0.721 to 0.735, suggesting that the spatial distribution radiomic features already capture most of the prognostic information that clinical stage and demographics provide.
For event prediction, the combined model achieved an AUC of 0.947 on the internal test and 0.782 on the external test, demonstrating strong classification ability to identify which patients would die during follow-up. The radiomics-only model also achieved an AUC of 0.909, underscoring the strong independent prognostic power of distribution-based features.
Feature selection retained 11 variables for the survival duration prediction model. These included two clinical features (age group and disease stage), the total TLG value, five TLGd radiomic features reflecting whole-body tumor spatial distribution, and three METAORG features from the axial skeleton, peripheral skeleton, and liver.
The 3D short-axis diameter of the combined tumor segment was one of the strongest prognostic features. In the context of whole-body tumor distribution, this metric reflects the distance between individual malignant lesions: patients with more widely spread disease (larger short-axis diameter across the body) had shorter survival, consistent with the known prognostic impact of widespread metastatic dissemination.
Greater variability in FDG uptake intensity across all tumor voxels and more heterogeneous spatial texture were associated with better prognosis in TLGd features, a counterintuitive finding that may reflect the presence of necrosis or treatment response within tumors. The authors note that further studies are needed to understand the biological basis of this relationship.
For organ-specific METAORG features, more homogeneous FDG uptake in the axial skeleton and more heterogeneous uptake in the peripheral skeleton were each associated with better survival, pointing to contrasting biological behaviors in central versus peripheral bone metastases. Uniform FDG uptake in the liver, reflecting the absence of focal nodular metastatic deposits, was also associated with better prognosis.
The prognostic advantage of tumor distribution over simple burden stems from the unique interpretation framework of METAORG. Most organs normally show homogeneous FDG uptake, so focal heterogeneities detected by organ-level radiomic texture features reflect the distributional pattern of metastases within that organ rather than simple aggregate metabolic activity.
The finding that skeletal and liver metastatic patterns were the strongest organ-specific predictors aligns with multiple meta-analyses identifying these sites as particularly harmful in SCLC. Quantifying how metastases are distributed within the skeleton and liver, not just whether they are present, provides richer stratification than binary presence or absence of involvement.
The METAORG approach sidesteps the labor-intensive and error-prone process of individually segmenting every metastatic lesion. By using automated CT organ segmentation transferred to PET, a single whole-organ VOI captures the collective metabolic signature of all metastases within that organ, potentially enabling automated and standardized prognostic assessment across institutions.
The authors propose that this approach could complement or improve upon RECIST criteria, which only assess tumor size changes, by incorporating distributional information that may better reflect biological heterogeneity in treatment response and could support treatment stratification for high-risk patients requiring more intensive or targeted therapy.
TotalSegmentator, a deep learning CT segmentation tool, automatically delineated 104 anatomic structures and was used to generate organ volumes of interest for METAORG analysis. Individual muscles and bones were grouped into clinically meaningful composite regions such as the axial skeleton and peripheral skeleton to capture metastatic distribution patterns at a biologically relevant scale.
CT-based organ volumes of interest were registered onto FDG PET images by leveraging the fact that CT and PET are acquired together in the same PET/CT session, meaning the CT used for attenuation correction defines organ boundaries that apply directly to the co-registered PET SUV data without additional registration steps.
Two types of METAORG features were extracted for each organ: total METAORG using the full organ VOI regardless of uptake level, and high METAORG using only voxels with SUV above 2.5 to isolate regions of definitely elevated metabolic activity. Both feature sets were included as inputs to the survival models.
This automated pipeline substantially reduces the clinical burden compared to manual tumor segmentation of all metastatic lesions, and by relying on CT-based organ boundaries rather than tumor-specific thresholds, it is more robust to variations in individual lesion conspicuity that could otherwise introduce segmentation errors.
The most immediate clinical application of the METAORG methodology is providing clinicians with a quantifiable metric for metastasis sub-stratification. Rather than categorizing patients simply as having or not having liver or bone metastases, the approach can generate numerical distributional scores that reflect the severity and pattern of organ involvement, enabling finer-grained risk stratification.
The approach could be applied as a semi-automatic screening tool that processes a whole-body PET/CT scan and outputs a composite prognostic score incorporating both tumor burden and organ distribution patterns, potentially integrated into a clinical nomogram for individualized survival probability estimation at the time of diagnosis.
Treatment response monitoring is another major potential application. Current RECIST-based response assessment only evaluates changes in the largest measurable lesions and does not capture redistributional changes such as new organ involvement or changes in within-organ metastasis patterns. METAORG could quantify these distributional changes across treatment cycles.
The authors note that the approach may extend beyond SCLC to any highly FDG-avid malignancy, including lymphoma and squamous cell cancers, where organ-level metabolic heterogeneity reflects tumor involvement patterns. However, it is not suitable for malignancies with inherently low FDG uptake such as thyroid or mucinous cancers.
This study demonstrates that how cancer spreads across the body and within specific organs is more prognostically important than how much cancer there is as measured by total metabolic burden. This finding has implications for how PET/CT data is interpreted and used in clinical decision-making for SCLC.
The combined model consistently outperformed all other combinations across both internal and external validation cohorts, and the external validation C-index of 0.740 suggests reasonable generalizability despite differences in patient demographics and treatment regimens between the two hospital cohorts.
The automated nature of the organ segmentation component makes this approach scalable to clinical implementation without requiring additional physician time for lesion-by-lesion delineation, addressing a key barrier to broader adoption of whole-body PET radiomic analysis.
Future prospective studies with larger cohorts, standardized scanner protocols, recurrence-site documentation, and inclusion of brain imaging and mediastinal lymph node segmentation would further validate and refine this approach toward routine clinical integration in SCLC management.