Endocrine Abstracts

JOINT3966

Thyroid cancer incidence has been increasing worldwide over the last decade. Early detection of nodules from volumetric imaging tests such as CT or MRI due to other medical problems account for a fraction of new cases, however, ultrasound imaging is the most commonly used detection tool. Confirmed nodules are scored for risk of malignancy based on the American College of Radiology (ACR) Thyroid Imaging Reporting & Data System (TI-RADS). TI-RADS scores are based on five features: composition, echogenicity, shape, margin and echogenic foci. Composition is scored from 0 to 2 points, while echogenicity, shape, margin and echogenic foci can receive up to 3 points. The sum of points and size of nodule are used to determine the need for a fine-needle aspiration (FNA) procedure to confirm malignancy. Automated scoring of ultrasound images based on convolutional neural networks (CNNs) have been proposed and can provide useful pointers to a radiologist, but remain essentially “black-box” models, failing to provide novel insights about the output they produce. In this work, we propose a CNN-based automated scoring tool trained on TI-RADS that learns a non-linear mapping from ultrasound image to a Euclidean space, which we refer to as the “disease space”. Inspired by Mihail et al., (Mihail, 2015) we impose the following constraints during the learning phase of the transformation to the new space: 1) L2 norm = TI-RADS score and 2) distance between similar images is minimized. Cluster analysis in this space reveals specific disease modes. Critically, this allows for nodule progression prediction by interpolating along the L2 norm vector. While the CNN component of our tool acts as a black-box, the disease space’s L2 norm one-to-one correspondence with TI-RADS score allows us to gain novel insights into malignancy prediction, given enough data points and ground truth from FNA. In our evaluation, we use 2 datasets: 347 thyroid ultrasound images of thyroid nodules from the DDTI (Digital Database Thyroid Image) and 13 thyroid ultrasound images of thyroid nodules from a private hospital in Romania. These patients signed an informed consent form committing to the use of their data only for scientific research.