<p>In clinical practice, the detection of key anatomical structures/organs in medical images is critical for a range of downstream tasks, such as quality control and diagnosis. However, deep learning-based models trained on medical images from one institution/device typically experience a performance decline when transferred directly to other institutions/devices due to domain gaps. Moreover, unlike natural images, medical images often contain significant noise and feature dense, overlapping structures, making the detection of anatomical structures more challenging. Thus, to tackle this problem, we propose a new Unsupervised Domain Adaptation (UDA) method integrated with the enhanced Anatomy-oriented Topology Knowledge Transfer (ATKT), Anatomy-aligned Morphology Knowledge Transfer (AMKT), and Auxiliary Network and Ensembling Learning (ANEL) modules for anatomical structure detection in medical images, named ToMo-UDA++. ATKT utilizes existing knowledge of human anatomy as enhanced topological information to reconstruct and align anatomical features across source and target domains. On the other hand, AMKT formulates a more consistent and independent enhanced morphological representation for each substructure of an organ. Then, we also devise an auxiliary network to supervise source domain training and generate high-quality pseudo-labels for the target domain respectively. Finally, we provide a comprehensive optimization of ATKT and AMKT, as well as alignment of the source and target domain anatomical structure features through construct templates, uncertainty-aware anatomical knowledge alignment, and bidirectional alignment. To fully evaluate the proposed ToMo-UDA++ for anatomical structure detection, we introduce one new benchmark, i.e., <Emphasis Type="ItalicUnderline">F</Emphasis>etal <Emphasis Type="ItalicUnderline">U</Emphasis>ltra<Emphasis Type="ItalicUnderline">S</Emphasis>ound <Emphasis Type="ItalicUnderline">S</Emphasis>pine dataset from <Emphasis Type="ItalicUnderline">T</Emphasis>hree <Emphasis Type="ItalicUnderline">D</Emphasis>evice with six annotated anatomical regions called <i>FUSSD</i><InlineEquation ID="IEq1"> <EquationSource Format="TEX">\(^3\)</EquationSource> <EquationSource Format="MATHML"><math> <mmultiscripts> <mrow /> <mrow /> <mn>3</mn> </mmultiscripts> </math></EquationSource> </InlineEquation>, which is the first dataset for pure cross-device multi-structure detection. 22 groups of adaptation experiments were conducted on the proposed benchmarks and publicly available datasets in various medical scenarios, i.e., cross-medical center, cross-modality, and cross-medical device. The results of the experiment show that the use of enhanced topological and morphological anatomy information in ToMo-UDA++ greatly improves organ/structure detection. ToMo-UDA++ and <i>FUSSD</i><InlineEquation ID="IEq2"> <EquationSource Format="TEX">\(^3\)</EquationSource> <EquationSource Format="MATHML"><math> <mmultiscripts> <mrow /> <mrow /> <mn>3</mn> </mmultiscripts> </math></EquationSource> </InlineEquation> dataset expand the potential for structure detection tasks in medical image analysis. Datasets and source codes are available at <a href="https://github.com/Yore0/ToMo-UDA_plus">https://github.com/Yore0/ToMo-UDA_plus</a>.</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

ToMo-UDA++: Unsupervised Domain Adaptation for Anatomical Structure Detection Using Enhanced Topology and Morphology Knowledge

  • Bin Pu,
  • Jiewen Yang,
  • Xingguo Lv,
  • Xingbo Dong,
  • Lei Zhao,
  • Shengli Li,
  • Kenli Li,
  • Xiaomeng Li

摘要

In clinical practice, the detection of key anatomical structures/organs in medical images is critical for a range of downstream tasks, such as quality control and diagnosis. However, deep learning-based models trained on medical images from one institution/device typically experience a performance decline when transferred directly to other institutions/devices due to domain gaps. Moreover, unlike natural images, medical images often contain significant noise and feature dense, overlapping structures, making the detection of anatomical structures more challenging. Thus, to tackle this problem, we propose a new Unsupervised Domain Adaptation (UDA) method integrated with the enhanced Anatomy-oriented Topology Knowledge Transfer (ATKT), Anatomy-aligned Morphology Knowledge Transfer (AMKT), and Auxiliary Network and Ensembling Learning (ANEL) modules for anatomical structure detection in medical images, named ToMo-UDA++. ATKT utilizes existing knowledge of human anatomy as enhanced topological information to reconstruct and align anatomical features across source and target domains. On the other hand, AMKT formulates a more consistent and independent enhanced morphological representation for each substructure of an organ. Then, we also devise an auxiliary network to supervise source domain training and generate high-quality pseudo-labels for the target domain respectively. Finally, we provide a comprehensive optimization of ATKT and AMKT, as well as alignment of the source and target domain anatomical structure features through construct templates, uncertainty-aware anatomical knowledge alignment, and bidirectional alignment. To fully evaluate the proposed ToMo-UDA++ for anatomical structure detection, we introduce one new benchmark, i.e., Fetal UltraSound Spine dataset from Three Device with six annotated anatomical regions called FUSSD \(^3\) 3 , which is the first dataset for pure cross-device multi-structure detection. 22 groups of adaptation experiments were conducted on the proposed benchmarks and publicly available datasets in various medical scenarios, i.e., cross-medical center, cross-modality, and cross-medical device. The results of the experiment show that the use of enhanced topological and morphological anatomy information in ToMo-UDA++ greatly improves organ/structure detection. ToMo-UDA++ and FUSSD \(^3\) 3 dataset expand the potential for structure detection tasks in medical image analysis. Datasets and source codes are available at https://github.com/Yore0/ToMo-UDA_plus.