Quantification of peripheral and central vascularization of thyroid nodules in power-doppler ultrasound 3D images

Standard Procedure for Thyroid Cancer Assessment Thyroid nodules have a very high incidence worldwide, but on average, approximately only 5% of them is malignant. The diagnostic procedure for the assessment of their nature is an ultrasound examination followed by fine needle aspiration biopsy (FNA). However, FNA is invasive, painful, complications may occur and usually only 9-13% of analysed nodules result in being malignant. These reasons show the need to reduce the number of unnecessary biopsies and increment the efficiency of feature extraction from US images. Study of Vascularization As pointed out in different studies, vascular pattern could be useful not only in avoiding FNA when not necessary, but also in the characterization of nodules with undetermined fine-needle aspiration biopsy results (THY 3a, THY 3b). Moreover, different studies suggest the existence of a correlation between a predominant central vascularity distribution and malignancy. Power Doppler is used to obtain information on the vascularization. In our work, volumetric PDUS acquisitions have been studied extracting architectural, tortuosity and flux parameters from the vascular network of 17 benign nodules (THY 3a) and 15 malignant nodules (THY 3b). Semi-automatic Segmentation of the Vascular Network First, a rough segmentation of the volume of the nodule was manually performed to prevent the extraction of parameters from structures that are not part of the nodule. Moreover, it is the necessary starting point to perform a distinction between peripheral and central vasculature of the nodule. Consequently, two volumetric masks have been created from the original mask to segment the peripheral and central volume of the nodule and extract the parameters. Specifically, erosion was performed on the original 3D mask and the central mask was produced. The 3D volume enclosing the peripheral circulation was obtained by subtraction between the total and central ones. Then, masks have been applied onto the 3D skeleton in order to extract architectural and tortuosity parameters and onto the 3D PDUS to extract flow parameters. They have been calculated both for the centre and the periphery of the nodule. Parameters Extraction The previous mentioned architectural parameters are: Number of Trees (NT), Vascular Ratio (VR) and Number of Branching points (NB). About the tortuosity metrics, Distance Metric (DM), Inflection Count Metric (ICM) and Sum of Angles Metric (SOAM) have been extracted. The same procedure was carried out in the extraction of flow parameters such as Vascular Fraction Area (A), Mean Intensity Index (VI) and Blood Flow Volume Index (FVI). Results and Discussion All the values of the parameters have been extracted for each patient. Pearson’s correlation analysis has been performed on them to discard the irrelevant and redundant ones. Firstly, the correlation of each variable with a target vector has been calculated and variables with a correlation value lower than a threshold have been discarded. Then the correlation between all the features has been calculated and, within the couples of features with a higher level of correlation, only the one with the higher correlation with the target has been kept. Classification learning with medium Gaussian SVM model has been performed on the relevant features, 8 group cross-validation has been chosen to validate the results. This classification showed an accuracy of 75% with sensitivity of 60 % and specificity of 88%.