Introduction
Segmenting lung tumors in computed tomography (CT) scans remains a significant challenge in medical imaging. Tumors exhibit extreme variability in size and shape, and their boundaries are frequently obscured by partial volume effects. Furthermore, aggressive tumors often invade adjacent anatomical structures, such as the chest wall or mediastinum, complicating standard segmentation approaches that rely on rigid lung masking.
Existing deep learning models face three primary hurdles: extreme class imbalance between small tumors and large CT volumes, rigid anatomical priors that may exclude invasive tumor regions, and a lack of uncertainty quantification, which limits clinical trust. This paper presents a two-stage framework designed to address these issues through coarse-to-fine refinement and anatomy-aware post-processing.
Methodology
The proposed framework utilizes a two-stage cascade pipeline to improve segmentation accuracy and reliability.
Stage 1: Coarse Localization and Anatomy-Aware Post-Processing
The first stage employs a 3D model (e.g., Swin UNETR) to process the full CT volume, acting as a surrogate for tumor localization. Rather than accepting this output as final, the system applies a "clinically informed" post-processing module:
- Component Merging: Fragmented predictions are joined using dilation.
- Anatomy-Aware Filtering: Instead of a binary lung mask, the system uses a nuanced rule: regions with low lung overlap are discarded unless they are in close proximity to the lung surface, allowing for the inclusion of invasive tumors.
- Component Selection: The authors found that selecting only the largest predicted component (Top-1) significantly reduces false positives, acting as a spatial attention mechanism.
Stage 2: Fine Refinement and Uncertainty Modeling
In the second stage, the model crops the CT image around the refined regions of interest (ROIs) for high-resolution segmentation. To handle boundary ambiguity, the model incorporates Monte Carlo (MC) Dropout to estimate voxel-wise uncertainty. An adaptive loss function dynamically reweights Dice and Cross-Entropy losses based on this uncertainty, forcing the network to focus on the most challenging boundary regions.
Results and Discussion
The framework was evaluated on the Orlando Health dataset and the public NSCLC-Radiomics dataset.
Performance Gains
The two-stage approach yielded substantial improvements. On the Orlando dataset, the Dice score for the Swin UNETR baseline increased from 0.4690 to 0.6447. Boundary accuracy also improved dramatically; the 95th percentile Hausdorff Distance (HD95) dropped from 187.97 mm in the full-volume baseline to 8.89 mm in the refined ROI model.
The Impact of Denoising
A key finding was the strong correlation between the reduction of spurious components and overall segmentation performance (ρ = -0.83, p < 0.0001). This validates the "anatomy-aware" filtering strategy as a critical component for denoising initial predictions. The "Top-1" selection strategy proved particularly effective at removing the noise that typically inhibits high Dice scores in single-stage models.
Conclusion
By treating segmentation as a coarse-to-fine process and integrating clinical priors with uncertainty modeling, this framework significantly enhances the precision of lung tumor delineation. While the "Top-1" strategy may require further refinement to handle multifocal tumors, the current system provides a robust, interpretable tool for clinical workflows, offering both improved accuracy and valuable uncertainty heatmaps for clinician review.
Summary updated