Cervical cancer is the leading cause of gynecologic disease-related death worldwide, but is almost completely preventable with regular screening, for which cytological testing is a method of choice. Although such testing has radically lowered the death rate from cervical cancer, it is plagued by low sensitivity and inter-observer variability. Moreover, its effectiveness is still restricted because the recognition of shape and morphology of nuclei is compromised by overlapping and clumped cells. Multispectral imaging can aid enhanced morphological characterization of cytological specimens. Features including spectral intensity and texture, reflecting relevant morphological differences between normal and abnormal cells, can be derived from cytopathology images and utilized in a detection/classification scheme. Our automated processing of multispectral image cubes yields nuclear objects which are subjected to classification facilitated by a library of spectral signatures obtained from normal and abnormal cells, as marked by experts. Clumps are processed separately with reduced set of signatures. Implementation of this method yields high rate of successful detection and classification of nuclei into predefined malignant and premalignant types and correlates well with those obtained by an expert. Our multispectral approach may have an impact on the diagnostic workflow of cytological tests. Abnormal cells can be automatically highlighted and quantified, thus objectivity and performance of the reading can be improved in a way which is currently unavailable in clinical setting.

Date Published: 9 March 2010
PDF: 10 pages
Proc. SPIE 7624, Medical Imaging 2010: Computer-Aided Diagnosis, 762435 (9 March 2010); doi: 10.1117/12.844482

Published in SPIE Proceedings Vol. 7624:
Medical Imaging 2010: Computer-Aided Diagnosis
Nico Karssemeijer; Ronald M. Summers, Editor(s)