Classification of thyroid nodules based on analysis of margin characteristic
Hanung Adi Nugroho, Eka Legya Frannita, Anan Nugroho, Zulfanahri, Igi Ardiyanto, Lina Choridah · 2017
The radiologists diagnose thyroid cancer by analysing thyroid ultrasound nodule images. However, it tends to be subjective since it depends on the expertise and experience of the radiologists. Therefore, a computer aided diagnosis (CADx) system is necessary to to reduce subjectivity and to support the radiologists in final decision making of thyroid cancer diagnosis. This study aims to classify thyroid nodules of ultrasound images by analysing margin characteristics. The proposed approach is evaluated on 144 images with 64 smooth and 80 irregular margins. Some noises and artefacts are eliminated by employing adaptive median filter and speckle reducing bilateral filtering (SRBF). Thyroid nodule is then segmented based on morphological operation and active contour. In order to classify segmented nodules, a total of eight geometric features are extracted and subsequently undergo classification process. Two different kernels of support vector machine (SVM) consisting of linear and quadratic kernels are used to evaluate the performance of classification. Evaluation results show that the quadratic kernel has better performance than the linear ones with the accuracy of 92.30%, sensitivity of 91.88%, specificity of 92.73%, PPV of 92.80% and NPV of 91.80%. These results indicate that the proposed approach successfully classifies thyroid nodule based on margin characteristics and is useful for assisting the radiologists in diagnosing thyroid cancer by analysing thyroid ultrasound images.