Transformer-based endoscopic polyp detection with structural reparameterization and dynamic upsampling

Abstract Accurate polyp detection is important for colorectal cancer prevention, while practical deployment requires reliable localization, manageable model complexity, and robust detection performance. This paper presents an improved RT-DETR-r18 detector for endoscopic polyp detection. The proposed C2f_RCB backbone construction embeds a UniRepLKNet-derived large-kernel reparameterization transformation within the C2f hierarchy, while the original DySample operator replaces the two nearest-neighbor upsampling layers in the top-down neck. A source-code audit shows that the K = 7 DilatedReparamBlock contains one 7 × 7 depthwise main branch and three auxiliary depthwise branches, namely 5 × 5 with dilation 1, 3 × 3 with dilation 2, and 3 × 3 with dilation 3. During deployment, these spatial branches are merged into an equivalent 7 × 7 depthwise convolution, and compatible normalization/affine terms are further folded into adjacent operators by the UniRepLKNetBlock deployment routine; SE, GELU, GRN normalization/scaling, channel projections, residual operations, and DySample remain explicit in the inference graph. Under the archived evaluation protocol, the complete model achieves 96.8% precision, 93.0% recall, 97.1% [email protected], and 79.5% [email protected]:0.95, with 13.90 million parameters and 46.6 GFLOPs. These values were obtained from an expanded-file split later found to contain augmented variants of the same original samples across subsets and should therefore be regarded as preliminary until leakage-free retraining is completed. The current evidence supports a compact, localization-oriented trade-off, but not a standardized runtime advantage, clinical readiness, or leakage-free generalization.