SegFormer with Hausdorff distance-based spatial bias and conceptual split gate for automated segmentation of cervical precancerous lesions

Automated segmentation of acetowhite lesions in Visual Inspection with Acetic Acid (VIA) cervicography is critical for colposcopy-guided biopsy targeting in low-resource cervical cancer screening programs. Standard transformer architectures compute attention solely from feature content, neglecting spatial proximity between tokens, while unified decoder designs conflate semantic and boundary channels—leaving boundary accuracy at the squamocolumnar junction unresolved.We propose SegFormer-HF-CSG, a parameter-efficient framework that addresses these limitations through two targeted modifications to the SegFormer backbone (as Illustrated in Figure 1):•Hausdorff Distance-based Spatial Bias (HF): injects pairwise spatial proximity information into the self-attention mechanism across all encoder stages, improving geometric coherence of acetowhite boundary predictions without adding trainable parameters.•Conceptual Split Gate (CSG): decouples the decoder into independent semantic and boundary pathways with adaptive soft gating, adding only 59,712 parameters (+2.3%).Evaluated on 545 VIA cervicography images, SegFormer-HF-CSG achieves DSC of 96.91%, HD95 of 3.46 mm, and Normalized Surface Distance of 98.82%, outperforming eight reference architectures spanning convolutional, hybrid, and transformer families (p < 0.01, Bonferroni-corrected) at a total of 2.62 M parameters—the smallest footprint in the evaluated set.