CMCNet: Aligning Ultrasound Embeddings with TI-RADS for Fine-Grained Thyroid Classification
A new study on arXiv (2608.13939) introduces CMCNet, a deep learning framework aimed at aligning ultrasound image embeddings with textual descriptions from TI-RADS to improve thyroid nodule classification. This research addresses the often overlooked area of multi-class predictions in thyroid ultrasound analysis, which has mainly focused on binary cancer detection. The team has created the STN dataset, containing 600 thyroid nodules, along with matching transverse and longitudinal ultrasound images, bounding box annotations, and detailed labels for all five TI-RADS feature categories. During training, the model leverages structured feature data for representation learning, relying solely on images for inference. The findings suggest that using text embeddings from standardized features can boost classification accuracy. This paper is categorized as a cross-type announcement and is available on arXiv.
Key facts
- CMCNet is a deep learning model for thyroid nodule classification.
- It aligns ultrasound image embeddings with textual TI-RADS representations.
- The STN dataset includes 600 thyroid nodules with paired images and annotations.
- The dataset has complete labels for all five TI-RADS feature categories.
- The model uses feature-level supervision during training but only images at inference.
- The study focuses on multi-class prediction, not just binary malignancy.
- The paper is available on arXiv with ID 2608.13939.
- The announcement type is cross.
Entities
Institutions
- arXiv