Application of a Feature Fusion-Based Improved Yoloseg Model in Thyroid Nodule Segmentation

Recently, the incidence of thyroid nodules and thyroid cancer has continued to rise, and early detection and timely intervention are crucial for preventing disease progression. Ultrasound, as a non-invasive imaging modality, has become the preferred method for nodule screening. However, due to the complex morphology and blurred borders of nodules, especially the identification of small lesions, traditional manual diagnosis is prone to subjective interference. With the development of artificial intelligence, this paper optimises and improves the YOLO11-seg model by introducing the BQA module in the segmentation head to enhance nodule detection capability. The GSH module introduces shared convolutional structures between multi-scale features and combines them with the scale layer to scale features, thereby reducing parameter overhead while improving detection accuracy for targets of different scales. Additionally, the EFA module is introduced in the C3K2 feature extraction part of the model to enhance the model's ability to perceive the edges of targets at different scales. Compared to the baseline model, our model reduces GFLOPs by approximately 11% while improving mAP50 by 9.6% and Precision by 8.8%, significantly enhancing detection accuracy while reducing computational costs.