Computer Aided Diagnosis: Initial Results for the Detection of Thyroid Nodules using US Images

Ziel/Aim Thyroid nodule detection and classification uses Ultrasound (US) imaging to provide the patient with an initial diagnosis of their condition. Detection and classification of a nodule is dependent on the experience of the physician and could be subjective. The aim of this study is to presentthe initial results obtained towards developing a Computer Aided Diagnostic (CAD) system for the detection of thyroid nodules in US images.

Methodik/Methods US scans were acquired prospectively and consecutively from 47 patients (Female = 76.6 %, Male = 23.4 %) with a total of 78 thyroid nodules. US videos were acquired while the scanning US probe was tracked within an electromagnetic field generated by an EM field generator. With that atotal of 2290 nodule images were extracted from the US video files with a mean nodule size of 14mmx12mmx16mm(width x depth x length). The extracted image dataset was subsequently used in a Convolutional Neural Network (CNN) to detect nodules in the ultrasound image. A CNN is a type of neural network used in image processing and recognition that is specifically designed to process pixel data (U-net Architecture). The dataset was split into a training set (70 %) and testing set (30 %).

Ergebnisse/Results The detection of the nodules was dependent on the size and composition of the nodules. Larger, cystic and hypoechogenic nodules showed a better distinction from normal thyroid tissue. This could be due to the difference in size and pixel intensity between the two regions. The network was able to detect the nodules with a training accuracy of 97.3 % and a validation accuracy of 95.0 % as compared to the clinically generated ground truth.

Schlussfolgerungen/Conclusions The current results demonstrateda high validation accuracy. Future work would be to further improve the network, test the trained and validated model on different image data sets and perform nodule texture classification. Unsupervised learning methods could significantly reduce the subjectivity that is still seen in the clinically generated ground truth.