MMTC-Net: Multimodal Temporal Cervical Network for HSIL+ Recognition in Precancer Screening

Cervical precancer screening is essential for reducing disease-related mortality. In colposcopic practice, clinicians jointly assess dynamic acetic-acid image sequences, iodine-stained images, and structured clinical information when distinguishing High-grade squamous intraepithelial lesion or higher (HSIL+) from low-grade squamous intraepithelial lesion or lower (LSIL-). However, most existing artificial intelligence models fail to integrate all three modalities effectively. To bridge this gap, we propose MMTC-Net, a lightweight multimodal temporal network for HSIL+ recognition in cervical precancer screening. MMTC-Net includes a temporal normalization linear module, which models frame-to-frame progression in acetic-acid sequences using a reference-normalized linear mapping, thereby preserving temporal cues with low computational overhead. MMTC-Net also uses a two-stage cross-modal attention mechanism that first aligns the acetic-acid and iodine-stained image modalities and then fuses the resulting image representation with structured clinical data. Evaluated by fivefold cross-validation on a real-world cohort of 1347 patients, MMTC-Net achieved a mean accuracy of 94.24%, sensitivity of 90.60%, specificity of 97.87%, and area under the receiver operating characteristic curve of 0.9765, significantly outperforming comparator models ( $p<0.05$ ). Ablation studies further confirmed the contributions of temporal modeling of acetic-acid image sequences and multimodal fusion of cervicograms and clinical data. By explicitly modeling the information sources used in routine colposcopic assessment, these results suggest that MMTC-Net can support reproducible AI-assisted decision-making, particularly in resource-constrained screening settings where experienced colposcopists are limited.