| 武涛,魏利胜,邵自强.基于轻量级改进RT-DETR的内窥镜息肉检测[J].电子测量与仪器学报,2025,39(11):246-257 |
| 基于轻量级改进RT-DETR的内窥镜息肉检测 |
| Endoscopic polyp detection based on lightweight improved RT-DETR |
| |
| DOI： |
| 中文关键词: 息肉检测 RT-DETR HiLo高低频分离机制 FasterNet |
| 英文关键词:polyp detection RT-DETR HiLo high-low frequency separation mechanism FasterNet |
| 基金项目:安徽省教育厅重大项目(KJ2020ZD39)、安徽省高等学校省级质量工程项目(2023cxtd057)资助 |
|
|
|
|
| 摘要点击次数: 414 |
| 全文下载次数: 378 |
| 中文摘要: |
| 针对息肉检测任务中存在息肉尺度差异显著、肠道环境复杂,以及医疗诊断设备资源有限影响检测精度的问题,提出一种基于RT-DETR(real-time detection transformer)改进的轻量级息肉检测模型。首先,采用FasterNet作为RT-DETR模型主干网络,重构FasterNet Block模块,分流冗余特征的同时提升对息肉的关注度;其次,设计了新模块,在内尺度特征交互(attention-based intrascale feature interaction, AIFI)内部引入HiLo（H-AIFI）高低频分离机制,分离局部高频细节和低频全局结构,聚焦复杂背景下的关键病灶点;最后,设计选择性边界聚合-特征金字塔网络（SBA-FPN）重校准特征融合网络替换跨尺度特征融合模块(cross-scale feature fusion module, CCFM),促进不同分辨率特征之间的双向融合,提升多尺度特征融合效果。实验结果表明,在公开的内窥镜息肉组合数据集上,与原始RT-DETR模型相比,改进模型mAP@0.5和mAP@0.5:0.95值分别提高2.3%和3.0%,参数量和计算量分别减少44.4%、48.6%。在Br35H脑肿瘤数据集上,改进模型mAP@0.5提高1.3%。由此可知,改进模型不仅满足息肉自动检测需求,而且满足医疗场景下泛化病灶的高精度检测。 |
| 英文摘要: |
| Aiming at the problems of significant differences in polyp size, complex intestinal environment, and limited medical diagnostic equipment resources affecting detection accuracy in polyp detection tasks, a lightweight polyp detection model based on RT-DETR improvement was proposed. Firstly, FasterNet is used as the backbone network of the RT-DETR model to reconstruct the FasterNet Block module to divert redundant features while increasing attention to polyps. Secondly, the new module is designed to introduce HiLo high and low frequency separation mechanism into the attention-based intrascale feature interaction (AIFI) to separate local high frequency details and low frequency global structures, and focus on key lesions in complex backgrounds. Finally, an SBA-FPN recalibration feature fusion network is designed to replace the cross-scale feature fusion module (CCFM) to promote two-way fusion between features with different resolutions and improve the multi-scale feature fusion effect. The experimental results show that compared with the original RT-DETR model, the mAP@0.5 and mAP@0.5:0.95 values of the improved model are increased by 2.3% and 3.0% respectively, and the amount of parameters and calculations is reduced by 44.4% and 48.6% respectively. On the Br35H brain tumor dataset, the mAP@0.5 of the improved model increased by 1.3%. It can be seen that the improved model not only meets the needs of automatic polyp detection, but also meets the high-precision detection of generalized lesions in medical scenarios. |
| 查看全文 查看/发表评论 下载PDF阅读器 |
|
|
|
