A Multi-Branch Feature Fusion Transformer Network and Its Application in Neck Ultrasound Detection.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Endocrinology

Background:

• Hyperparathyroidism (HPT) and thyroid nodules (TN) originate from parathyroid and thyroid glands, respectively.
• Their anatomical proximity and similar ultrasound features complicate accurate differentiation in medical images.
• Traditional object detection algorithms face challenges in distinguishing coexisting HPT and TN, leading to potential misdiagnosis.

Purpose of the Study:

• To develop an accurate object detection method for differentiating coexisting hyperparathyroidism and thyroid nodules in ultrasound images.
• To improve diagnostic accuracy for patients with concurrent parathyroid and thyroid conditions.
• To address the limitations of existing algorithms in complex ultrasound scenarios.

Main Methods:

• Construction of three specialized datasets: HPTD (hyperparathyroidism only), TND (thyroid nodules only), and HPT-TND (mixed lesions).
• Development of a novel Multi-Branch Feature Fusion DETR network (MB-DETR), based on the RT-DETR model.
• Integration of redesigned feature fusion modules and asymmetric convolution for enhanced feature extraction.

Main Results:

• The MB-DETR model demonstrated superior performance over state-of-the-art models in F1, Precision, and Recall metrics.
• Significant reduction in computational costs was achieved compared to existing methods.
• Ablation studies validated the effectiveness of asymmetric convolution and multi-branch feature fusion in improving detection.

Conclusions:

• The proposed MB-DETR model effectively enhances local feature extraction capabilities within the DETR framework.
• MB-DETR outperforms current models in detecting coexisting thyroid nodules and hyperparathyroidism.
• This AI-driven approach offers significant potential for assisting in the clinical diagnosis of related endocrine diseases.