Integrating Fisheye Transformation and Multi-View Voting for Improved Lesion Localization In Chest CT
Abstract
Purpose
Automated segmentation of lung nodules in chest CT is critical for early cancer screening but remains challenging due to the small size and variable morphology of nodules, which often resemble vessels or pleura. This study proposes a novel framework integrating fisheye-inspired spherical magnification with multi-center view fusion to enhance local lesion visibility and decision robustness using an nnU-Netv2 baseline.
Methods
A total of 840 chest CT cases from the LIDC-IDRI dataset were employed (700 for training, 140 for testing). To address the challenge of small targets, we implemented a 3D fisheye spherical deformation strategy, which non-linearly magnifies the local 3D region around a target center while compressing the periphery. For each case, 16 magnification centers were uniformly sampled on a 2×2×4 grid within the lung region (defined via TotalSegmentator). These centers generated 16 locally magnified views per case to train a standard 3D nnU-Netv2. During inference, predictions from the 16 views were fused using a lesion-level voting scheme, requiring detection agreement from at least 8 views to retain a candidate. Finally, lesion-level precision, recall, and F1-score were evaluated, with a 0.3 cc volume threshold applied to exclude clinically insignificant tiny lesions.
Results
On the 140-case test set, the baseline nnU-Netv2 achieved a lesion-based recall of 0.870, precision of 0.780, and F1-score of 0.820. The proposed framework significantly improved performance, reaching a recall of 0.872, precision of 0.893, and F1-score of 0.883. These results suggest fewer false positives without sacrificing lesion detection.
Conclusion
The proposed fisheye-based spherical magnification and multi-center voting strategy effectively suppresses false positives without compromising sensitivity. By enhancing local feature visibility and enforcing multi-view consistency, this approach offers a robust solution for detecting subtle lung nodules and is easily integrable into existing 3D segmentation architectures.