BLUE RIBBON POSTER MULTI-DISCIPLINARY: Radiomics-Driven Volumetric Active Surface Optimization (R-VASO): Overcoming Low-Contrast Segmentation Failures In Deep Learning Auto-Contouring
Abstract
Purpose
Deep Learning (DL) auto-segmentation models rely on voxel-wise probability maps, often resulting in jagged boundaries or "spill-over". Traditional post-processing, such as 2D active contours, fails in low-contrast scenarios due to the lack of edge gradients. This study proposes a novel 3D post-processing framework, Radiomics-Driven Volumetric Active Surface Optimization (R-VASO), which uses organ-specific textural gradients rather than intensity gradients to drive contour convergence, constrained by 3D minimal surface energy to ensure volumetric topological consistency.
Methods
We used a dataset of 20 abdominal CT images with expert-verified contours. DL segmentations were generated using a U-Net architecture. The R-VASO framework was applied in two stages, first radiomic potential field is generated for each target organ, where a specific radiomic signature (e.g., GLCM Entropy for liver) was identified that maximized the separability between the organ and adjacent soft tissue, effectively creating a "texture contrast" where "intensity contrast" was absent. A gradient vector flow field was computed from these feature maps. Secondly DL contours were converted into a 3D triangular mesh. This mesh will evolve under the influence of the Radiomic Potential Field (external force) and a Mean Curvature Flow constraint (internal force) to minimized the total surface energy, effectively regularizing z-axis inconsistencies.
Results
The R-VASO framework improved contour accuracy compared to standard active contour refinement in low contrast regions. The Dice Similarity Coefficient improved from 0.84±0.06 (raw DL) to 0.91±0.03 (R-VASO). The 95% Hausdorff Distance was reduced by an average of 1.8mm (p < 0.01). The 3D surface energy constraint eliminated "z-flicker" — discontinuous shapes between adjacent axial slices—resulting in smooth, clinically acceptable volumes with much less manual intervention.
Conclusion
R-VASO improves segmentation accuracy by combining texture-based driving force with 3D topological regularization. This method provides a robust solution for refining AI-generated contours in low-contrast situation, potentially reducing manual editing time by 40%.