<p>Pulmonary nodule detection is essential for the early diagnosis and prevention of lung cancer. Existing studies have proposed various enhancement algorithms tailored to pulmonary nodules, but small-sized nodules are still prone to false detections, making accurate detection challenging. To address these limitations and enhance the accuracy of lung nodule detection, we propose an end-to-end framework referred to as GD-RTDETR, which primarily includes a newly designed cross-scale feature-fusion backbone network. Specifically, a cross-scale fusion mechanism is embedded into the backbone to enhance the capture of fine-grained features from small nodules. In addition, a Cascaded Group Attention (CGA) module is incorporated to avoid information isolation of small nodule features and to improve the efficiency of intra-scale feature interaction. Furthermore, a Gated Sampling and Dual Fusion Enhanced Path Aggregation Network (GDPAN) is proposed, which dynamically extracts informative features by embedding weight information into the feature maps and performs efficient multiscale fusion, thereby improving the detection of small-sized nodules. The experimental results show that GD-RTDETR increases AP, <InlineEquation ID="IEq1"> <InlineMediaObject> <ImageObject Color="BlackWhite" FileRef="40747_2025_2032_Article_IEq1.gif" Format="GIF" Height="16" Rendition="HTML" Resolution="72" Type="Linedraw" Width="37" /> </InlineMediaObject> <EquationSource Format="TEX">\(\hbox {AP}_{50}\)</EquationSource> <EquationSource Format="MATHML"><math> <msub> <mtext>AP</mtext> <mn>50</mn> </msub> </math></EquationSource> </InlineEquation>, <InlineEquation ID="IEq2"> <InlineMediaObject> <ImageObject Color="BlackWhite" FileRef="40747_2025_2032_Article_IEq2.gif" Format="GIF" Height="16" Rendition="HTML" Resolution="72" Type="Linedraw" Width="37" /> </InlineMediaObject> <EquationSource Format="TEX">\(\hbox {AP}_{75}\)</EquationSource> <EquationSource Format="MATHML"><math> <msub> <mtext>AP</mtext> <mn>75</mn> </msub> </math></EquationSource> </InlineEquation>, and <InlineEquation ID="IEq3"> <InlineMediaObject> <ImageObject Color="BlackWhite" FileRef="40747_2025_2032_Article_IEq3.gif" Format="GIF" Height="16" Rendition="HTML" Resolution="72" Type="Linedraw" Width="34" /> </InlineMediaObject> <EquationSource Format="TEX">\(\hbox {AP}_{S}\)</EquationSource> <EquationSource Format="MATHML"><math> <msub> <mtext>AP</mtext> <mi>S</mi> </msub> </math></EquationSource> </InlineEquation> by 4.1%, 1.0%, 8.1%, and 4.0%, respectively, achieving the best performance among all the models compared.</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

Enhanced pulmonary nodule detection using a transformer framework with dual fusion and gated mechanism

  • Kelei Sun,
  • Yihang Wang,
  • Huaping Zhou

摘要

Pulmonary nodule detection is essential for the early diagnosis and prevention of lung cancer. Existing studies have proposed various enhancement algorithms tailored to pulmonary nodules, but small-sized nodules are still prone to false detections, making accurate detection challenging. To address these limitations and enhance the accuracy of lung nodule detection, we propose an end-to-end framework referred to as GD-RTDETR, which primarily includes a newly designed cross-scale feature-fusion backbone network. Specifically, a cross-scale fusion mechanism is embedded into the backbone to enhance the capture of fine-grained features from small nodules. In addition, a Cascaded Group Attention (CGA) module is incorporated to avoid information isolation of small nodule features and to improve the efficiency of intra-scale feature interaction. Furthermore, a Gated Sampling and Dual Fusion Enhanced Path Aggregation Network (GDPAN) is proposed, which dynamically extracts informative features by embedding weight information into the feature maps and performs efficient multiscale fusion, thereby improving the detection of small-sized nodules. The experimental results show that GD-RTDETR increases AP, \(\hbox {AP}_{50}\) AP 50 , \(\hbox {AP}_{75}\) AP 75 , and \(\hbox {AP}_{S}\) AP S by 4.1%, 1.0%, 8.1%, and 4.0%, respectively, achieving the best performance among all the models compared.