Semantic Image Analysis for Cervical Neoplasia Detection

This paper presents an automated semantic image analysis method for cervical cancerous lesion detection. We model colposcopic image semantics in a novel probabilistic manner using conditional random fields. We extract the anatomical structure of the cervix from colposcopic images, and identify and summarize different tissue types and their locations in an image semantics map. The conditional random field model uses the semantic information to model the unique optical properties of each tissue type and the diagnostic relationships between neighboring regions. We validate our method using clinical data from 48 patients, and the results demonstrate its diagnostic potential in detecting neoplastic areas. Our automated diagnostic approach has the potential to support or substitute for conventional colposcopy. Furthermore, our generalized framework can be applied to other cancers such as skin, oral and colon cancer.