Scholarlink

<p>CT scans offer a valuable way to automatically segment knots, however wet logs pose a crucial challenge for sawmills, due to the similar intensity between water saturated sapwood and the knot. Most existing approaches rely on traditional segmentation techniques or 2D surface level imaging limiting their effectiveness for internal knot characterisation. To address this, we propose a 3D segmentation model tailored specifically for Norway spruce knots, using a comprehensive dataset of 24 trees from three different regions of Sweden (384 CT volumes). We systematically benchmarked popular state of the art 3D deep learning architectures used in medical image segmentation utilising the MONAI framework. Our new model KnotSegNet3D combines residual blocks in both encoder and decoder for improved boundary delineation. The models were evaluated using wet stem sections, reflecting sawmill conditions, while training involved a mixture of training on dry and wet data along with transfer learning from dry to wet. Mixed training achieved the highest accuracy (Dice: <InlineEquation ID="IEq1"> <EquationSource Format="TEX">\({\textbf {0.902}} \pm {\textbf {0.025}}\)</EquationSource> </InlineEquation>) and the closest boundaries (HD95: <InlineEquation ID="IEq2"> <EquationSource Format="TEX">\({\textbf {1.610}} \pm {\textbf {0.190}}\)</EquationSource> </InlineEquation> mm), higher than the strongest baseline Attention UNet model (Dice: <InlineEquation ID="IEq3"> <EquationSource Format="TEX">\({\textbf {0.786}} \pm {\textbf {0.115}}\)</EquationSource> </InlineEquation>; HD95: <InlineEquation ID="IEq4"> <EquationSource Format="TEX">\({\textbf {6.57}} \pm {\textbf {12.47}}\)</EquationSource> </InlineEquation> mm), despite having fewer parameters in KnotSegNet3D (15.26M vs 23.63M). Transfer learning also showed similar performance: Full fine-tuning reached Dice <InlineEquation ID="IEq5"> <EquationSource Format="TEX">\({\textbf {0.896}}\pm {\textbf {0.029}}\)</EquationSource> </InlineEquation> and HD95 <InlineEquation ID="IEq6"> <EquationSource Format="TEX">\({\textbf {1.620}}\pm {\textbf {0.180}},\textrm{mm}\)</EquationSource> </InlineEquation>, close to mixed training (<InlineEquation ID="IEq7"> <EquationSource Format="TEX">\({\textbf {0.902}} \pm {\textbf {0.025}}\)</EquationSource> </InlineEquation>; <InlineEquation ID="IEq8"> <EquationSource Format="TEX">\({\textbf {1.610}}\pm {\textbf {0.190}},\textrm{mm}\)</EquationSource> </InlineEquation>). In addition, partial fine-tuning reduced training time per epoch due to fewer trainable parameters, providing a practical option when computational resources or annotations are limited. Qualitatively, KnotSegNet3D handled complex knot geometries and consistently delivered better precision and stability in terms of performance index and visual results for trees selected from different regions.</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

KnotSegNet3D: high precision knot segmentation in Norway spruce CT logs via mixed training and fine-tuned transfer learning

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年4月24日

查看全文
  • Mohammad Jaber Hossain,
  • Geir Isak Vestøl,
  • Olof Broman,
  • Thomas Reichert,
  • Oliver Tomic,
  • Linus Olofsson

摘要

CT scans offer a valuable way to automatically segment knots, however wet logs pose a crucial challenge for sawmills, due to the similar intensity between water saturated sapwood and the knot. Most existing approaches rely on traditional segmentation techniques or 2D surface level imaging limiting their effectiveness for internal knot characterisation. To address this, we propose a 3D segmentation model tailored specifically for Norway spruce knots, using a comprehensive dataset of 24 trees from three different regions of Sweden (384 CT volumes). We systematically benchmarked popular state of the art 3D deep learning architectures used in medical image segmentation utilising the MONAI framework. Our new model KnotSegNet3D combines residual blocks in both encoder and decoder for improved boundary delineation. The models were evaluated using wet stem sections, reflecting sawmill conditions, while training involved a mixture of training on dry and wet data along with transfer learning from dry to wet. Mixed training achieved the highest accuracy (Dice: \({\textbf {0.902}} \pm {\textbf {0.025}}\) ) and the closest boundaries (HD95: \({\textbf {1.610}} \pm {\textbf {0.190}}\) mm), higher than the strongest baseline Attention UNet model (Dice: \({\textbf {0.786}} \pm {\textbf {0.115}}\) ; HD95: \({\textbf {6.57}} \pm {\textbf {12.47}}\) mm), despite having fewer parameters in KnotSegNet3D (15.26M vs 23.63M). Transfer learning also showed similar performance: Full fine-tuning reached Dice \({\textbf {0.896}}\pm {\textbf {0.029}}\) and HD95 \({\textbf {1.620}}\pm {\textbf {0.180}},\textrm{mm}\) , close to mixed training ( \({\textbf {0.902}} \pm {\textbf {0.025}}\) ; \({\textbf {1.610}}\pm {\textbf {0.190}},\textrm{mm}\) ). In addition, partial fine-tuning reduced training time per epoch due to fewer trainable parameters, providing a practical option when computational resources or annotations are limited. Qualitatively, KnotSegNet3D handled complex knot geometries and consistently delivered better precision and stability in terms of performance index and visual results for trees selected from different regions.