DD-CMD: Dual-domain text-guided segmentation hits 91.46% Dice accuracy

Clinical text can guide medical image segmentation, but existing text-guided models focus on spatial alignment and ignore frequency content that defines texture and boundaries. In a paper accepted at MICCAI 2026, researchers Md Maklachur Rahman and Tracy Hammond introduce DD-CMD (Dual-Domain Cross-Modal Decoding), a framework that integrates both spatial and frequency-domain language guidance. The decoder uses Text-Guided Spatial Cross-Attention (TGSA) to align multi-scale visual tokens with text semantics via gated residual fusion, alongside Spectral-Text Adaptive Modulation (STAM), which applies a 2D discrete cosine transform (DCT) to compute learnable band-energy statistics and predict text-conditioned FiLM parameters. This coarse-to-fine decoder progresses from 7x7 to 56x56 resolution and restores full-resolution masks with a lightweight two-stage refinement module.

DD-CMD was evaluated on two public pulmonary infection datasets: QaTa-COV19 and MosMedData+. It achieved 91.46% Dice / 84.26% mIoU on QaTa-COV19 and 81.95% Dice / 69.42% mIoU on MosMedData+, with average gains of +1.96 Dice and +2.67 mIoU over the strongest prior baselines. The authors highlight that frequency-domain modulation is key to capturing texture and boundary details that spatial-only methods miss. The code has been released publicly on GitHub. For clinicians and researchers working on automated COVID-19 assessment, DD-CMD offers a more accurate way to turn natural-language clinical notes into precise segmentation masks, potentially reducing manual annotation effort and enabling more reliable AI-assisted diagnosis.