Attention U-Net Segmentation of Indeterminate Nodules On Thyroid Ultrasound
J Genender*1, J Fuhrman1, H Li1, Kelvin Memeh1, J Conn Busch1, J Williams1, L Lan1, D Sarne1, B Finnerty2, P Angelos1, T Fahey2, X Keutgen1, M Giger1 (1) University of Chicago, Chicago, IL, (2) Weill Cornell Medicine, New York, NY
Presentations
PO-GePV-I-100 (Sunday, 7/10/2022) [Eastern Time (GMT-4)]
ePoster Forums
Purpose: Approximately 20-30% of thyroid biopsies return indeterminate diagnoses, often necessitating further testing or diagnostic surgeries. Recent reports note the capacity of artificial intelligence (AI) to distinguish benign from malignant thyroid nodules. However, these studies often use manual nodule segmentation, which is burdensome. Here we present our thyroid nodule segmentation method using a 2D U-Net with additive attention (Attention U-Net), a component of our overall AI ultrasound nodule characterization algorithm.
Methods: Our dataset consisted of 1,342 grayscale ultrasound images of 379 biopsy-proven thyroid nodules from 331 patients at Institution A and 48 patients at Institution B. These included 197 benign, 137 malignant, and 45 indeterminate nodules. The nodules were contoured by two experienced physicians to yield the segmentation truth (reference standard). U-Net models with additive attention underwent training/validation/testing (75/10/15) by nodule (i.e., not by image) on three types of input: original ultrasound images, regions of interest (ROIs) without resizing, and ROIs with resizing to 256x256 pixels. During training, Dice loss and binary cross-entropy loss functions with Adam optimization were used. Segmentation performances were assessed using the Sorensen-Dice coefficient (DSC).
Results: The average DSC of the predicted contours on the original images was 0.595 (SE=0.012). For unaltered ROIs and resized ROIs, the average DSCs of the predicted contours were 0.854 (SE=0.002) and 0.923 (SE=0.001), respectively. After Bonferroni corrections for multiple comparisons, all paired-sample T-tests indicated statistically significant differences (p<0.001) between the predicted contours of the original images and unaltered ROIs; original images and resized ROIs; and unaltered ROIs and resized ROIs.
Conclusion: The U-Net segmentation method delineated thyroid nodules within ultrasound images across three input sizes with varying effectiveness, achieving high segmentation accuracies for resized ROIs. This segmentation method is now being incorporated into our ultrasound AI system for computerized thyroid nodule delineation and diagnosis for rapid analysis.
Funding Support, Disclosures, and Conflict of Interest: This research was funded in part by the NIH S10 OD025081 Shared Instrument Grant and the AAPM Summer Undergraduate Fellowship Program.
Keywords
Segmentation, Ultrasonics, CAD
Taxonomy
IM/TH- Image Segmentation Techniques: Modality: Ultrasound
Contact Email