Patch Based Texture Classification of Thyroid Ultrasound Images using Convolutional Neural Network.

Ultrasound (US) is an affordable and important imaging modality in medical imaging without potential hazards for patients and medical practitioners as compared to computed tomography which uses X-rays, magnetic resonance imaging which uses magnetic field and radio waves that could heat up the patient's body during long examinations, nuclear imaging, etc. Texture classification of anatomical structures in US images is an essential step for disease diagnosis and monitoring. In this work, we employed a convolutional neural network to segment thyroid gland in US images. This is particularly important for thyroid diseases diagnosis as they involve changes in the shape and size of the thyroid over time. The training of the Convolutional Neural Network (CNN) was not done directly on the acquired US images but on texture database that is created by dividing the thyroid US images of size 760 × 500 pixels into smaller texture patches of size 20 × 20 pixels. We obtained a Dice coefficient (DC) of 0.876 and Hausdorff Distance (HD) of 7.3 using the trained CNN that classifies the thyroid tissues as thyroid or non-thyroid. This approach was compared to the classic image processing approaches like active contours with edges (ACWE), graph cut (GC) and pixel-based classifier (PBC) which obtained a DC of 0.805, 0.745 and 0.666 respectively and Volumetric and Mass-Spring Models which obtained a HD of 11.1 and 9.8 respectively.