Identification of Malignant Patterns in FNAC Digital Images of Thyroid Nodules through Cascaded Segmentation Stages

In this paper, an optimal computer aided diagnosis system with a cascade combination of two region-based segmentation stage is proposed and evaluated to discriminate benign thyroid nodules from malignant using thyroid Fine Needle Aspiration Cytology (FNAC) microscopic images. Two region-based image segmentation methods, namely, watershed and mathematical morphology are sequentially ensemble, in the first pass, to extract the foreground cell portions from thyroid FNAC images by discarding background staining information. Because of the ensembled sequential application of two segmentation methods, the majority of the unwanted background image pixels are eliminated, and the required foreground cells are retained in the images. In the second pass, the statistical features are taken by Wavelet decomposition and Gabor filter models. Finally, the Support Vector Machine (SVM) classifier is implemented on derived feature set for differentiating benign thyroid nodules from malignant. A promising 93.33% of highest diagnostic accuracy is obtained using SVM model for Gabor features which is 3.33% higher than the previous results. The new configuration suggests that the proposed system with cascade operation of two region-based segmentation methods can improve the performance for classifying multi-stained FNAC thyroid images as benign or malignant nodule.