Learning Multispectral Texture Features for Cervical Cancer Detection

Abstract

We present a bottom-up approach for automatic cancer cell detection in multispectral microscopic thin Pap smear images. Around 4,000 multispectral texture features are explored for cancer cell detection. Using two feature screening measures, the initial feature set is effectively reduced to a computationally manageable size. Based on pixel-level screening results, cancerous regions can thus be detected through a relatively simple procedure. Our experiments have demonstrated the potential of both multispectral and texture information to serve as valuable complementary cues to traditional detection methods.