Cross‐Slice Context Modeling via Hybrid Mamba‐ <scp>CNN</scp> for Universal Lesion Detection in <scp>CT</scp> Scans

ABSTRACT Universal lesion detection (ULD) in computed tomography (CT) scans is crucial for cancer diagnosis and staging. However, detecting lesions based on a single‐slice input is challenging due to their small size, low contrast, and the complexity of surrounding anatomical structures. As lesions often span multiple adjacent slices, capturing cross‐slice contextual information is crucial for developing automated lesion detection algorithms. To address this, we propose HyMamba‐ULD, a Hybrid Mamba‐CNN backbone designed for cross‐slice context modeling. Our approach consists of two key components: (1) the Target Slice Enhanced Mamba‐CNN (TSE‐MC) module, which extracts multiscale 3D context features from multiple adjacent slices by integrating local CNN‐based representations with global Mamba‐based features, and further refines them through an attention mechanism centered on the target slice; and (2) the Cross‐Slice Feature Aggregation Mamba (CSFAM) module, which explicitly incorporates global information from upper and lower slices to enrich the 2D feature representation of the target slice, thereby facilitating subsequent 2D accurate detection. We conduct evaluation on the NIH DeepLesion benchmark for the ULD task. Our approach attains a leading average recall of 89.24% across the range of 0.5–4 false positives per image. In addition, experiments on the LiTS dataset further verify the generalization and effectiveness of the proposed method. Source code is accessible via https://github.com/ssli23/HyMamba‐ULD .