Discrimination of Malignancy in Multi-Stained Thyroid FNAB Cytological Images

Abstract

Pattern classification problem in medical image analysis is normally solved under three phases: Segmentation of region of interest, Feature extraction and Classification. The classification accuracy is mainly depending on the efficient segmentation methodology. In our work, we have tested watershed transform and mathematical morphology based segmentation methods to segment the medullary and papillary carcinoma cell regions in multi-stained Thyroid Fine Needle Aspiration Biopsy (FNAB) cytological images. The performance of the segmentation methods has been evaluated by discriminating medullary and papillary carcinoma malignancies. Initially image segmentation is performed to remove the background information and retain the appropriate foreground cancer cell information in microscopic images. Feature extraction and classification are carried out by Discrete Wavelet Transform (DWT) and k-Nearest Neighbor (kNN) classifier respectively. The average classification rate of 92.22% is achieved with the combination of morphology segmentation method and kNN classifier whereas the combination watershed segmentation method and kNN classifier gives the average classification rate of 85.56%.