Small nodules frequently escape detection. Lung cancer screening with low-dose CT can reduce cause-specific mortality by approximately 20% compared to standard chest X-rays, but the benefit depends entirely on detecting nodules accurately. Lesions only a few millimeters in diameter may share similar attenuation values with surrounding lung tissue, or lie adjacent to blood vessels and pleural surfaces that obscure their margins.
Human review has real limits. Reader fatigue from high scan volumes, motion artifacts near the lung bases, and the sheer variability of nodule appearances contribute to missed detections during manual review. These practical constraints make reliable, automated computer-assisted detection an important clinical need rather than a luxury.
Classical algorithms have persistent false positive problems. Earlier automated detection systems often achieved high sensitivity but generated more than two false positive detections per scan - a burden that creates unnecessary follow-up work for radiologists and can lead to avoidable invasive procedures for patients. Reducing false positives while maintaining sensitivity is the central engineering challenge.
The LIDC-IDRI dataset as a benchmark. This study validates its pipeline against the Lung Image Database Consortium and Image Database Resource Initiative (LIDC-IDRI) collection, which contains over 1,000 expert-annotated thoracic CT studies. This publicly available dataset allows objective, reproducible comparison of competing detection methods.
Pipeline overview. The proposed system transforms raw CT scan volumes into precise three-dimensional nodule masks through four integrated stages: contrast enhancement, edge-preserving denoising, three-dimensional region growing from seed points, and geometric shape-based filtering to remove false positives. Each stage is specifically designed to address a known failure mode of earlier approaches.
Dual contrast stretching for visibility. CT values span over 3,000 Hounsfield units, but pulmonary tissue falls within a narrow window of roughly 400 units. Global intensity normalization followed by slice-wise percentile stretching (between the 2nd and 98th percentiles of each axial slice) expands the visible range for subtle nodules without amplifying noise - lifting faint nodules by approximately 10 grey levels compared to background tissue.
Edge-preserving anisotropic diffusion for denoising. A mathematical filter based on the Perona-Malik diffusion equation reduces noise while protecting nodule boundaries. The filter averages pixel values with neighboring voxels - but exponentially reduces this averaging wherever steep intensity gradients exist (i.e., at tissue boundaries). Ten iterations with carefully calibrated settings reduced slice-level noise variance by about 25% while preserving edge contrast within 5% of original values.
Adaptive seeding and three-dimensional region growing. Rather than using a single global intensity threshold to identify nodule candidates, the system computes local mean and standard deviation in a 20-voxel neighborhood around each point. A voxel becomes a seed if its value exceeds the local mean by 0.8 standard deviations. Seeds then expand in three dimensions, accepting neighboring voxels whose density differs from the growing region's mean by no more than 45 Hounsfield units - a tolerance calibrated on 95% of observed intra-nodule intensity variation.
Morphological repair and shape-based pruning. After region growing, binary closing fills internal holes and fractured boundaries, while binary opening removes thin protrusions. Two geometric criteria then filter the remaining objects: sphericity must be at least 0.65 (retaining compact, ball-like shapes) and eccentricity must not exceed 0.40 (eliminating elongated vascular segments). Together these rules eliminate 82% of non-nodule segments while reducing recall of true nodules by less than 3%.
Validated against expert annotations. Averaged across the full LIDC-IDRI cohort, the pipeline achieved a mean Dice overlap coefficient of 0.834, a sensitivity of 0.923, and a specificity of 0.987. The false positive burden was 1.46 detections per scan - below the clinical target of 2 per scan. Processing time for a 500-slice study was 154 seconds on standard CPU hardware without requiring a GPU.
Outperformed both baseline methods. Traditional threshold-based segmentation and watershed algorithms achieved Dice scores of 0.721 and 0.765 respectively, each generating more than 3 false positives per scan. The proposed method outperformed both on every metric, with particularly large gains on part-solid nodules smaller than 8 mm - the most clinically challenging subtype.
Accuracy increases with nodule size. Stratifying results by nodule size reveals an expected pattern: small nodules (3-6 mm) achieved a mean Dice of 0.80, medium nodules (6-10 mm) achieved 0.86, and large nodules (greater than 10 mm) achieved 0.90. Critically, the proposed method outperformed baselines by 0.12 Dice points for small nodules - the group where detection matters most and is hardest to achieve.
Consistency across scanner settings. Specificity remained nearly flat when reconstructing images with increasingly sharp processing kernels (from B31 to B80), dropping only from 0.988 to 0.985 - a range well within acceptable clinical tolerance. Both baseline algorithms showed steeper declines, indicating that edge-enhancing kernels introduced texture artifacts that they incorrectly flagged as nodules.
Variance-bounded region growing as the core innovation. Unlike fixed-threshold segmentation, the variance-bounded region growing approach continuously updates the region's mean intensity as each new voxel is accepted. This allows the growing front to adapt to gradual intensity changes within heterogeneous part-solid nodules, substantially reducing leakage into connected vessels.
Dual geometric gates for false positive removal. Combining sphericity and eccentricity thresholds as complementary filters addresses two distinct false positive sources: sphericity distinguishes compact nodule shapes from pleural tags (irregular shapes), while eccentricity removes elongated vascular segments. Using both together eliminated 82% of spurious detections while preserving 97% of true nodule recall.
Unsharp masking to recover edge detail. Anisotropic diffusion reduces noise but unavoidably softens some boundary sharpness. A final unsharp masking step - subtracting a slightly blurred version from the denoised image and adding back a scaled edge term - restored edge gradient magnitude from 13.4 to 22.1 Hounsfield units per pixel along nodule rims, a 65% improvement, with less than 2% reduction in overall signal-to-noise ratio.
Lightweight CPU-only implementation. The entire pipeline runs without a GPU, completing a 500-slice study in under 3 minutes and using less than 800 megabytes of memory even in 12-thread mode. This design ensures compatibility with standard clinical workstation hardware and avoids the IT infrastructure requirements of GPU-accelerated systems.
Thick-slice reconstructions cause missed detections. The most common cause of missed nodules in failure analysis was scan reconstruction with 5 mm slice thickness. At this resolution, partial volume averaging blurs nodule boundaries to the point where percentile stretching cannot restore sufficient contrast for reliable detection. Clinical programs using thinner slices (1-1.5 mm) would see better performance.
Cavitary lesions defeat the segmentation approach. Thin-walled cavitary nodules - those with hollow air-filled centers and thin tissue walls - present a specific failure mode. The low CT density inside the cavity biases local intensity statistics downward, causing the adaptive threshold to drop below the thin wall's grey level and halting region growth prematurely. This limitation motivates future work on edge-completion modules that enforce geometric priors on partially segmented cavities.
Robustness to increased noise. Simulated noise at progressively higher levels (adding Gaussian noise of 0-50 Hounsfield units) showed approximately linear Dice degradation, losing about 4 percentage points per 10 HU added. Even at the highest noise level, Dice remained at 0.70 - suggesting the pipeline can tolerate reduced tube current or older scanner acquisitions without catastrophic failure.
Anisotropic diffusion is the computational bottleneck. Runtime profiling showed that edge-preserving diffusion consumed 37% of total processing time, followed by neighborhood statistics (31%) and surface mesh extraction (14%). GPU acceleration of the diffusion step is the most impactful optimization target, potentially reducing total processing time from 154 seconds to under 30 seconds per study.
Ready to serve as a screening support tool. By achieving 0.834 Dice overlap, 92.3% sensitivity, and 1.46 false positives per scan within a 154-second runtime on commodity hardware, the pipeline meets the pre-specified targets set by the research team. These results suggest the system could meaningfully reduce the proportion of clinically relevant nodules missed during routine screening review.
Foundation for downstream AI classification. The high-fidelity 3D masks generated by this pipeline can serve as the structural backbone for subsequent deep learning classifiers that distinguish benign from suspicious nodules using radiomic texture signatures or transformer-based feature extraction. Separating the detection and classification steps - each optimized independently - is a design principle that may yield better overall performance than end-to-end approaches.
Planned future extensions. Several enhancements are identified: adding radiomic texture classifiers to stratify detected nodules by malignancy risk; enabling volumetric growth tracking by registering baseline and follow-up scans; porting to GPU to achieve under-30-second processing; and incorporating emphysema and air trapping scoring to broaden clinical utility beyond nodule management alone.
Reproducibility as a design principle. The paper provides fully explicit algorithmic parameters, pseudocode, and MATLAB-compatible implementations, enabling independent replication of all results. The use of the publicly available LIDC-IDRI benchmark for all validation ensures that comparison with future methods remains straightforward and transparent.