ThyGraph: A Graph-Based Approach for Thyroid Nodule Diagnosis from Ultrasound Studies

Improved thyroid nodule risk stratification from ultrasound (US) can mitigate overdiagnosis and unnecessary biopsies. Previous studies often train deep learning models using manually selected single US frames; these approaches deviate from clinical practice where physicians utilize multiple image views for diagnosis. This paper introduces ThyGraph, a novel graph-based approach that improves feature aggregation and correlates anatomically proximate images, by leveraging spatial information to model US image studies as patient-level graphs. Graph convolutional networks are trained on image-based and patch-based graphs generated from 505 US image studies to predict nodule malignancy. Self-attention graph pooling is introduced to produce a node-level interpretability metric that is visualized downstream to identify important inputs. Our best performing model demonstrated an AUROC of 0.866 ± 0.019 and AUPRC of 0.749±0.043 across five-fold cross validation, significantly outperforming two previously published attention-based feature aggregation networks. These previous studies fail to account for spatial dependencies by modeling images within a study as independent, uncorrelated instances. In the proposed graph paradigm, ThyGraph can effectively aggregate information across views of a nodule and take advantage of inter-image dependencies to improve nodule risk stratification, leading to better patient triaging and reducing reliance on biopsies. Code is available at https://github.com/ashwath-radha/ThyGraph.