LOCUS-Diff: Location-Controlled Thyroid Ultrasound Synthesis for Advancing Nodule Detection.
Authors
Affiliations (5)
Affiliations (5)
- College of Control Science and Engineering, Zhejiang University, Hangzhou, China.
- Department of Head and Neck Surgery, Sir Run Run Shaw Hospital, Zhejiang University School of Medicine, Hangzhou, China.
- Department of Ultrasound in Medicine, Sir Run Run Shaw Hospital, Zhejiang University School of Medicine, Hangzhou, China.
- Department of Ultrasound, Longyou People's Hospital, Quzhou, China.
- The School of Computer Science and Technology, Zhejiang University of Water Resources and Electric Power, Hangzhou, China. [email protected].
Abstract
Thyroid cancer is the most prevalent endocrine malignancy, and its diagnosis relies heavily on ultrasound for diagnosis and treatment planning. Although deep learning holds promise for automating this ultrasound diagnosis, its performance is severely bottlenecked by data scarcity. However, constructing high-quality datasets remains challenging due to strict privacy regulations, high annotation costs, and the inherent rarity of malignant samples. To address these challenges, we propose LOCUS-Diff, an innovative generative framework engineered for thyroid ultrasound synthesis targeting object detection, which integrates three essential modules: (1) a thyroid ultrasound synthesis foundation model capturing intricate acoustic textures and pathological semantics for clinically authentic synthesis; (2) a thyroid nodule spatial control branch that provides stable spatial control through explicit geometric constraints; and (3) a novel COMB (Correction Of Misaligned Boxes) mechanism designed to adaptively calibrate labels, thereby eliminating label noise. Rigorous visual Turing tests by senior clinical experts confirm that LOCUS-Diff generates anatomically plausible samples that rival real scans and outperforms other advanced methods in visual quality. Furthermore, extensive experiments on the TN5000 and TN3k datasets demonstrate that LOCUS-Diff consistently outperforms the state-of-the-art approach in downstream detection tasks. Notably, augmenting a training subset containing only 60% of the real data with synthetic samples yields a higher mAP than training on the full real dataset, highlighting improved data efficiency.