.png){kind=icon}
.png){kind=icon}
Return
TN-MFFD: a real-time end-to-end multi-scale feature fusion detector for thyroid nodules in ultrasound images
.png){kind=icon}
PRE
AI
S
Y
Y
J
DOI：10.1007/s00530-026-02599-x.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
The automatic detection of thyroid nodules in ultrasound images is challenged by drastic scale variations, speckle noise, and highly similar features between benign and malignant nodules. In this study, a real-time end-to-end detection framework named Thyroid Nodule Multi-scale Feature Fusion Detector (TN-MFFD) was proposed and applied to detect and perform class-aware nodule detection for thyroid nodules in ultrasound images. TN-MFFD is realized through a novel hybrid encoder integrating three purpose-designed modules: the Multi-scale Information Gathering and Distribution (MIGD) for cross-scale contextual reasoning, the Multi-scale Channel Fusion Attention (MCFA) for the reciprocal exchange of semantics and spatial details between adjacent feature scales, and the Fourier Injection (FInject) for frequency-domain global texture encoding. Extensive experiments were conducted on two distinct large-scale datasets, TN5000 and TN3K. The results demonstrate that TN-MFFD yields state-of-the-art performance, achieving an $$mAP_{50}$$ of 87.5% on TN5000 and 70.1% on TN3K. This represents improvements of 1.9% and 2.6% over the direct baseline DEIM and surpasses 13 other competing detection methods. Evaluated under a single-frame inference setting, TN-MFFD maintains visually smooth real-time monitoring. This demonstrates that our framework successfully balances high-precision diagnostic capability with the stringent real-time requirements of clinical ultrasound screening.
Keywords:
Deep learning
Object detection
Thyroid nodules
Ultrasound imaging
Journal
IF:
3.1
Papers:
2.7K
Citations:
2.7K
Organization
Cited Papers