Mamba-powered attention for auxiliary diagnosis of cervical cancer

Cervical cancer is the fourth most common malignant tumor among women worldwide, with particularly severe morbidity and mortality rates in areas with limited medical resources. A severe shortage of pathologists leads to insufficient screening coverage and low diagnostic efficiency. Deep learning technology offers new possibilities for computer-aided diagnosis; however, existing convolutional networks have limited receptive fields, making it difficult to model long-range spatial dependencies in billion-pixel full-slice images. While Vision Transformer possess global modeling capabilities, they impose an unbearable computational burden. More critically, existing methods cannot effectively combine local cellular features with global tissue information, neglecting the bidirectional information flow inherent in cervical pathology diagnosis. Based on this, this paper proposes a cervical cancer auxiliary diagnostic model, MPA-ADC, based on Mamba and attention mechanisms. This model comprises a segmentation network MANet and a dual-path classification network DPNet: MANet achieves efficient global feature capture through parallel Mamba modules and is enhanced by a multi-scale attention mechanism to improve feature fusion capabilities; DPNet extracts subtle features at the cellular level and structural information at the tissue level through independent local and global paths, respectively. Unlike existing works that simply stack modules, MPA-ADC formalizes the multi-scale feature coordination problem of pathological images into a dual-path state-space evolution problem and constructs an end-to-end solution with a state-space model as the mathematical kernel and an attention mechanism as the routing protocol. Experiments on the cervical cancer dataset show that this method achieves good performance in both cell nucleus segmentation and lesion classification tasks.