Morphology semantics-guided vision language alignment for cervical cell image classification

Cervical cytology screening is crucial for the early detection of cervical cancer and frequently leverages deep learning-based cell image classification. End-to-end models tend to rely on staining variations and statistical image features, which is inconsistent with clinical diagnostic principles with morphological criteria such as the nuclear-to-cytoplasmic ratio and nuclear membrane morphology. We propose a morphology semantics-guided framework for cervical cell image classification. We map discrete class labels to class-level morphological text descriptions according to the Bethesda System (TBS), and adopt an N-to-1 correspondence paradigm between image instances and class description. Then, we design a Class-aware Morphological Vision-Language Alignment (CaM-VLA) module that leverages bidirectional multi-granular cross-modal attention and class-aware contrastive learning to establish fine-grained correspondences between morphological semantics and image features. We further propose a similarity-based cross-modal matching scheme for classification, which is consistent with the working principle of CaM-VLA. In-domain and cross-class experiments show that explicitly incorporating morphology semantics improves classification performance. Qualitative visualization results show that the model may associate diagnostically relevant morphological concepts with image regions.