Patch-based classification of thyroid nodules in ultrasound images using direction independent features extracted by two-threshold binary decomposition

• Published in: Computerized medical imaging and graphics 2019-01, Vol.71, p.9-18
• Main Authors: Prochazka, Antonin, Gulati, Sumeet, Holinka, Stepan, Smutek, Daniel
• Format: Article
• Language: English
• Subjects: Accuracy; Benign; Classification; Computer-aided diagnosis; Decomposition; Diagnosis; Diagnosis, Computer-Assisted - methods; Diagnosis, Differential; Diagnostic systems; Feature extraction; Female; Humans; Image acquisition; Image classification; Image Interpretation, Computer-Assisted - methods; Image processing; Image segmentation; Inclination angle; Male; Medical imaging; Middle Aged; Nodules; Patch-based image analysis; Random forests; Sensitivity and Specificity; Support Vector Machine; Support vector machines; System effectiveness; Texture analysis; Thyroid; Thyroid gland; Thyroid Nodule - diagnostic imaging; Thyroid Nodule - pathology; Thyroid nodules; Ultrasonic imaging; Ultrasonography - methods; Ultrasound; Ultrasound imaging
• ISSN: 0895-6111; 1879-0771
• DOI: 10.1016/j.compmedimag.2018.10.001

•Random Forests were used for the patch-based classification of thyroid nodules and its performance was compared with the SVM classifier.•Random Forests classify malignant and benign nodules with ROC AUC = 0.971 when using all patches and ROC AUC = 0.999 when using 75% of all patches.•A new patch-based features extraction technique based on Two-Threshold Binary Decomposition is used for diagnosis of thyroid nodules.•Data from two different ultrasound devices was used in the study.•Patch-based classification operates with small squares of thyroid image measuring just 17 × 17 px, providing promisingly accurate results. Ultrasound imaging of the thyroid gland is considered to be the best diagnostic choice for evaluating thyroid nodules in early stages, since it has been marked as cost-effective, non-invasive and risk-free. Computer aided diagnosis (CAD) systems can offer a second opinion to radiologists, thereby increasing the overall diagnostic accuracy of ultrasound imaging. Although current CAD systems exhibit promising results, their use in clinical practice is limited. Some of the main limitations are that the majority use direction dependent features so, they are only compatible with static images in just one plane (axial or longitudinal), requiring precise segmentation of a nodule. Our intention has been to design a CAD system which will use only direction independent features i.e., not dependent upon the orientation or inclination angle of the ultrasound probe when acquiring the image. In this study, 60 thyroid nodules (20 malignant, 40 benign) were divided into small patches of 17 × 17 pixels, which were then used to extract several direction independent features by employing Two-Threshold Binary Decomposition, a method that decomposes an image into the set of binary images. The features were then used in Random Forests (RF) and Support Vector Machine (SVM) classifiers to categorize nodules into malignant and benign classes. Classification was evaluated using group 10-fold cross-validation method. Performance on individual patches was then averaged to classify whole nodules with the following results: overall accuracy, sensitivity, specificity and area under receiver operating characteristics (ROC) curve: 95%, 95%, 95%, 0.971 for RF and; 91.6%, 95%, 90%, 0.965 for SVM respectively. The patch-based CAD system we present can provide support to radiologists in their current diagnosis of thyroid nodules, whereby it can increase the overall accuracy of ul