Thyro-LMD: A Benchmark Dataset and Sample-Driven Data Loading, Attention, and Regularization for Long-Tailed...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Pathology

Background:

• Computer-aided diagnostic (CAD) methods for thyroid ultrasound (TUS) face challenges with real-world diagnostic guidelines that involve complex lesion descriptor combinations.
• Existing methods struggle with the long-tailed distributions inherent in these combinations, limiting robustness and generalizability.

Purpose of the Study:

• Introduce Thyro-LMD, the first long-tailed multi-label dataset for TUS, annotated with ACR TI-RADS lexicons using histopathology as reference.
• Benchmark existing AI methods, including large multimodal models, on this dataset to identify performance gaps, particularly for underrepresented classes.
• Propose SynTUS-Net, a novel CAD baseline designed to address long-tailed multi-label classification challenges in TUS.

Main Methods:

• Developed Thyro-LMD, a dataset with fine-grained annotations reflecting real-world TUS label distributions.
• Benchmarked various AI models (end-to-end, multimodal large models, foundation models) on Thyro-LMD.
• Designed and implemented SynTUS-Net, a purpose-built network with specialized modules for data loading, feature encoding, and prediction regularization to handle data imbalance.

Main Results:

• Thyro-LMD exhibits a highly imbalanced label distribution, challenging existing AI methods.
• Representative methods showed limitations in classifying tail-class lesions.
• SynTUS-Net achieved state-of-the-art performance on Thyro-LMD, outperforming conventional models and GPT-4o, especially on tail classes (e.g., 42.76% improvement in Tail-F1 over GPT-4o).

Conclusions:

• Thyro-LMD and SynTUS-Net provide a clinically grounded benchmark and a new approach for interpretable and generalizable AI in TUS.
• The proposed SynTUS-Net effectively addresses the long-tailed multi-label classification problem in medical imaging.
• This work paves the way for more robust AI-driven diagnostic tools in ultrasound imaging.