Integrating Fisheye Transformation and Multi-View Voting for Improved Lesion Localization in Chest CT

Abstract

AbstractAutomated detection and segmentation of pulmonary nodules in chest computed to- mography (CT) are essential for early lung cancer screening and computer-aided diagnosis. However, accurate identification of small nodules remains challenging because they often occupy only a small fraction of the CT volume and may resemble normal anatomical struc- tures such as vessels or pleural attachments. These characteristics frequently lead to missed detections or excessive false positives in deep learning–based segmentation systems. This dissertation investigates strategies for improving lesion-level detection reliability by enhanc- ing local lesion visibility and introducing multi-view decision consistency. This study proposes a framework that integrates a fisheye-inspired spherical magnifica- tion transformation with multi-center view fusion based on a strong 3D nnU-Netv2 base- line. The fisheye transformation applies a radially decaying nonlinear coordinate remapping that increases sampling density around a deformation center while preserving surrounding anatomical context. For each CT case, sixteen deformation centers are generated within the lung region using a uniform 2 ˆ2 ˆ4 spatial grid, producing sixteen locally magnified views that are used during training and inference. During inference, predictions from all views are combined using a lesion-level voting mechanism that requires agreement from at least eight views, together with a small-volume filtering step to remove unstable predictions. Experiments were conducted on the publicly available LIDC-IDRI chest CT dataset containing 840 cases (700 for training and 140 for testing). Evaluation was performed using lesion-level, case-averaged metrics including recall, precision, and F1-score. Compared with a standard 3D nnU-Netv2 baseline, the proposed approach achieved substantially improved precision and F1-score while maintaining comparable recall. The results demonstrate that combining fisheye-based local magnification with multi-view voting effectively suppresses false positives without compromising detection sensitivity. This framework provides a prac- tical and architecture-independent strategy for improving the reliability of small-lesion de- tection in medical image segmentation systems.