Scholarlink

<p>The detection of benign and malignant endometrial lesions is of paramount importance in clinical settings. However, due to the high heterogeneity and dynamic nature of endometrial tissue, existing detection methods fail to simultaneously meet the dual requirements of high precision and rapid detection. To address these challenges, we designed a lightweight real-time detection model, HD-YOLO, to improve detection performance in complex medical imaging. First, we propose Dual-Domain Fusion Convolution (DDFConv), a novel module that combines time- and frequency-domain convolutions to capture structural features of lesion areas from multiple perspectives. By leveraging complementary information from dual domains, it enhances the model’s feature extraction performance compared to traditional single-domain convolution. Second, we propose Multi-Scale Channel Fusion (MSCF), an innovative module that integrates convolutional kernels of different sizes. This enables the network to learn multi-scale features, thereby enhancing the model’s perception capabilities. Furthermore, we incorporated wavelet pooling into the down-sampling process. By integrating low-frequency and high-frequency information, this approach preserves more image details and texture information. Compared to YOLO11, HD-YOLO achieves a 16.7% reduction in parameters, a 12.7% reduction in GFLOPs, and a 14.8% reduction in model size. Meanwhile, it improves mAP50 by 1.4% and FPS by 13.9%, reaching an mAP50 of 79.5%. These innovations not only advance lightweight detection models but also provide a robust theoretical basis for the precise diagnosis of gynecological diseases.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

HD-YOLO: real-time detection of benign and malignant endometrial lesions using YOLO11

  • 期刊论文

  • 发表日期：

    2025年5月12日

查看全文
  • Xi Chen

摘要

The detection of benign and malignant endometrial lesions is of paramount importance in clinical settings. However, due to the high heterogeneity and dynamic nature of endometrial tissue, existing detection methods fail to simultaneously meet the dual requirements of high precision and rapid detection. To address these challenges, we designed a lightweight real-time detection model, HD-YOLO, to improve detection performance in complex medical imaging. First, we propose Dual-Domain Fusion Convolution (DDFConv), a novel module that combines time- and frequency-domain convolutions to capture structural features of lesion areas from multiple perspectives. By leveraging complementary information from dual domains, it enhances the model’s feature extraction performance compared to traditional single-domain convolution. Second, we propose Multi-Scale Channel Fusion (MSCF), an innovative module that integrates convolutional kernels of different sizes. This enables the network to learn multi-scale features, thereby enhancing the model’s perception capabilities. Furthermore, we incorporated wavelet pooling into the down-sampling process. By integrating low-frequency and high-frequency information, this approach preserves more image details and texture information. Compared to YOLO11, HD-YOLO achieves a 16.7% reduction in parameters, a 12.7% reduction in GFLOPs, and a 14.8% reduction in model size. Meanwhile, it improves mAP50 by 1.4% and FPS by 13.9%, reaching an mAP50 of 79.5%. These innovations not only advance lightweight detection models but also provide a robust theoretical basis for the precise diagnosis of gynecological diseases.