LOCUS-Diff: Location-Controlled Thyroid Ultrasound Synthesis for Advancing Nodule Detection

Proposes LOCUS-Diff, a thyroid ultrasound synthesis framework targeting object detection, integrating three modules: a synthesis foundation model capturing acoustic textures and pathological semantics, a nodule spatial control branch providing stable spatial control through explicit geometric constraints, and a COMB (Correction Of Misaligned Boxes) mechanism that adaptively calibrates labels. Relative to prior generative approaches, it brings spatial controllability and label-noise correction into a single synthesis pipeline aimed at downstream detection rather than image realism alone. The framework is described as a modular design, and COMB is stated to eliminate label noise; implementation details are not expanded in the provided summary-level text.

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. Uses clinical expert assessment as evidence of synthesis quality rather than relying only on automated metrics. Evidence comes from visual Turing tests by senior clinical experts, which are expert subjective judgments; the text does not report test size or statistics.

Extensive experiments on the TN5000 and TN3k datasets show that LOCUS-Diff consistently outperforms the state-of-the-art approach in downstream detection tasks. Translates synthesis quality into improved downstream detection performance rather than image-level improvement alone. Based on experiments on two named datasets (TN5000, TN3k), reported as consistently outperforming the state of the art, though specific numbers are not listed.

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. Indicates that synthetic samples can maintain or improve detection performance while reducing the need for real data. Uses mAP as the comparison metric, a quantitative result; the text does not provide specific mAP values or confidence intervals.

This work targets the specific setting of thyroid ultrasound nodule detection, suited to medical imaging research environments where privacy regulations, annotation costs, and the rarity of malignant samples make large high-quality datasets difficult to build; its value lies in offering a synthetic augmentation path for training detection models under data scarcity and a modular approach that other detection tasks requiring spatially controllable synthesis can draw on.

The provided text is summary-level and lacks specific mAP values, the number of experts and criteria in the visual Turing tests, the size and splits of TN5000 and TN3k, and the algorithmic details of COMB; these affect judgment of synthesis quality and the magnitude of data-efficiency gains and warrant consulting the original figures and experimental setup. In addition, how synthetic samples generalize in real clinical workflows remains to be further observed.