Answers are grounded in this paper and cite their sources.
Structured thyroid ultrasound diagnosis requires simultaneous recognition of multiple sonographic attributes and their clinically competing options. However, most deep learning methods treat these attributes as independent prediction tasks, limiting their ability to model diagnostic hierarchy, evidence reliability, and option-level competition. In this paper, we propose DERCNet, a diagnosis-guided evidence reasoning network for structured thyroid ultrasound diagnosis. DERCNet introduces three key components. First, Rule-Aware Hierarchical Clinical Option Evidence Prompting (RA-HCOEP) constructs clinically constrained option representations by integrating attribute membership, option semantics, risk priors, and image-adaptive context. Second, Reliability-Guided Evidence Routing (RGER) extracts multi-granularity ultrasound evidence and routes reliable evidence slots to corresponding clinical options, enabling option-level visual evidence assignment without dense evidence annotations. Third, Clinical Attribute Reliability-Aware Evidence Counterfactual Competition (CARE-CC) compares each target option with reliability-weighted sibling evidence within the same attribute, enforcing clinically meaningful competition among mutually exclusive diagnostic candidates. Experiments on the private WZThyroid dataset and two public datasets, DDTI and TNSCUI, demonstrate the effectiveness of DERCNet for structured diagnosis and its transferability to external binary thyroid nodule classification. On WZThyroid, DERCNet achieved 88.48% average accuracy. On the public DDTI and TNSCUI datasets, it improved the best baseline accuracy by 3.57 and 2.13 percentage points, respectively. These results indicate that modeling structured thyroid ultrasound diagnosis as clinical option-level evidence reasoning, rather than independent attribute prediction, provides a more balanced and clinically aligned solution.
No takes yet. Share an insight, caveat, or question.
娄恒凯 et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: