Lymphovascular invasion (LVI) is a critical determinant of prognosis in colorectal cancer (CRC), reflecting a tumour’s capacity for dissemination through blood and lymphatic vessels. Accurate preoperative assessment of LVI could inform treatment decisions and improve outcomes, yet remains limited by the need for postoperative histological confirmation. Traditional imaging modalities such as CT and PET/CT contribute to staging and evaluation, but their standard analyses lack the granularity to detect LVI reliably. Radiomics has offered a non-invasive route for characterising tumour heterogeneity, although its assumption of uniformity within tumours limits its accuracy. A novel method based on habitat analysis of 18F-fluorodeoxyglucose (FDG) PET images may overcome these limitations by segmenting tumours into biologically distinct subregions and quantifying their spatial interactions.
Habitat Analysis for Intratumoural Heterogeneity
The habitat-based approach divides tumours into multiple subregions based on voxel-level radiomic characteristics, offering a detailed view of intratumoural heterogeneity. Using FDG-PET/CT scans from 177 CRC patients, researchers calculated a new index—the intratumoural heterogeneity score (ITHscore)—by clustering voxel data into three distinct habitats and assessing spatial relationships through multiregional spatial interaction metrics. This process captured subtle variations within tumour microenvironments and integrated both tumoural and peritumoural features. ITHscore was derived using a random forest classifier trained with recursive feature elimination and optimised with a balanced dataset. In validation, the ITHscore demonstrated a strong association with LVI status, outperforming conventional radiomic analysis in stability and generalisation.
Comparative Performance of Predictive Models
The predictive capacity of four models was assessed: a Clinical model using PET/CT-reported lymph node status; a Radscore model based on standard radiomics; an ITHscore model derived from habitat analysis; and a Combined model incorporating all three elements. In the training cohort, the Combined model achieved the highest area under the curve (AUC: 0.857), with strong accuracy, sensitivity, and specificity. Its performance remained consistent in the validation cohort (AUC: 0.798), significantly surpassing the Clinical and Radscore models. Although the ITHscore model alone also performed well (AUC: 0.712), its integration within the Combined model led to superior classification and clinical utility, as evidenced by decision curve analysis. Importantly, the ITHscore showed no significant correlation with Radscore, indicating it captures distinct and complementary tumour information.
Must Read: Predicting Liver Metastasis in Rectal Cancer with Radiomics
Clinical Relevance and Implementation Considerations
Understanding the presence of LVI preoperatively has major implications for CRC management, particularly for planning surgical margins and adjuvant therapy. The ITHscore, as a stable and non-invasive indicator, offers a new dimension of risk stratification. Traditional metabolic PET parameters such as SUVmax, metabolic tumour volume (MTV), and total lesion glycolysis (TLG) did not independently predict LVI status, underscoring the added value of heterogeneity-focused analysis. Moreover, the nomogram developed from the Combined model aligned closely with observed outcomes in both training and validation cohorts, supporting its practical utility. The method’s reliance on standard PET/CT imaging makes it broadly applicable, though further prospective and multicentre validation is needed to confirm its generalisability and clinical integration.
Quantifying intratumoural heterogeneity through habitat analysis of 18F-FDG PET images provides a valuable tool for preoperative prediction of LVI in colorectal cancer. The ITHscore offers improved insight into tumour biology beyond standard radiomics, and its integration with clinical and imaging variables enhances predictive accuracy. As a non-invasive and reproducible biomarker, it holds potential to inform personalised treatment strategies and optimise outcomes for CRC patients. Further research and automation may pave the way for widespread clinical adoption.
Source: Academic Radiology
Image Credit: iStock
