Abstract: With additional independent multisite validation, COLLAGE measurements might enable noninvasive characterization of the degree of recurrent tumor or cerebral radiation necrosis in gadolinium-contrast T1WI of posttreatment lesions.
Introduction
Distinguishing between recurrent tumor (RT) and cerebral radiation necrosis (CRN) is a critical diagnostic challenge in neuro-oncology. Both conditions appear as enhancing lesions on standard post-treatment T1-weighted MRI due to blood-brain barrier disruption. However, their clinical management differs drastically: RT requires aggressive intervention, while CRN is typically managed conservatively. Currently, definitive differentiation often necessitates invasive brain biopsy.
Methodology: The COLLAGE Framework
This study proposes a "virtual biopsy" using COLLAGE (Co-occurrence of Local Anisotropic Gradient Orientations), a radiomic feature that quantifies the "disorder" of pixel-level edge orientations. Unlike traditional intensity-based metrics, COLLAGE captures the structural heterogeneity of the tissue microenvironment.
The researchers analyzed a multi-institutional cohort of 75 gadolinium-contrast T1-weighted MRI studies, including primary and metastatic tumors from three different sites and the Cancer Imaging Archive (TCIA). The algorithm calculates image gradients, generates entropy maps to measure local disorder, and extracts statistical features (mean, median, skewness, kurtosis) for classification via a Random Forest model.
Key Results and Insights
The study revealed that recurrent tumors are biologically more heterogeneous and structurally disorganized than necrotic tissue, leading to higher entropy in MRI pixel gradients.
Skewness vs. Signal Intensity
A pivotal finding was that standard MRI signal intensity failed to distinguish between tumor and necrosis ($P > .05$). In contrast, skewness—a measure of the asymmetry of the entropy distribution—emerged as a highly significant differentiator. Skewness values were significantly lower for pure CRN ($0.15 \pm 0.12$) compared to predominant recurrent tumors ($0.42 \pm 0.21$).
Performance on Mixed Pathologies
The COLLAGE system demonstrated the ability to track the degree of necrosis. In "mixed" cases (e.g., 50% tumor / 50% necrosis), the feature values fell between the pure categories, validating that the metric correlates with the underlying pathological spectrum. The classification model achieved accuracies of 68.2% for primary tumors and 65% for metastatic tumors.
Conclusion
These preliminary findings suggest that COLLAGE-based entropy mapping can non-invasively differentiate RT from CRN by identifying subtle structural patterns invisible to the human eye. While limited by sample size and its retrospective nature, this approach offers a promising path toward reducing the need for invasive biopsies in post-treatment brain tumor monitoring. Future work will focus on larger prospective trials and integrating multi-sequence MRI data.
Summary updated