Scholarlink

Abstract <p>Accurate intraoperative identification of bone and cartilage is critical for orthopedic surgery, particularly during joint replacement procedures. Current approaches rely on direct visual inspection, which may be compromised by limited visibility or tissue deformation. This study represents the first investigation of hyperspectral imaging (HSI) for orthopedic surgery, assessing its potential for real-time intraoperative tissue segmentation when combined with lightweight artificial intelligence (AI) models, evaluated offline on in vivo porcine data. We leveraged the publicly available HeiPorSPECTRAL dataset, selecting Bone, Cartilage, and an (<i>Others</i>) class as representative orthopedic tissues. Pixel-based (multilayer perceptron (MLP), one-dimensional convolutional neural network (1D-CNN)) and patch-based (two-/three-dimensional U-Net (2D/3D U-Net)) models were systematically compared, with and without spectral normalization. Performance was assessed using accuracy, Dice similarity coefficient, and inference time. Normalization degraded performance across all models. The MLP achieved the best trade-off between accuracy (97.03%) and computational efficiency, segmenting a full hyperspectral cube in <InlineEquation ID="IEq1"> <EquationSource Format="TEX">\(&lt;1\)</EquationSource> </InlineEquation> s. While 3D U-Net provided slightly higher Dice scores (0.96-0.99), inference required &#xa0;8 s per cube, limiting clinical applicability. Qualitative inspection confirmed the robustness of the MLP in delineating bone and cartilage, even in challenging intraoperative scenarios. HSI combined with lightweight AI enables accurate and rapid tissue segmentation, highlighting its potential as a real-time intraoperative guidance tool in orthopedic surgery. Future work will focus on spectral channel reduction, in vivo validation, and integration with surgical instrumentation to accelerate clinical translation. We provide the first evidence that HSI can be applied to orthopedic tissue segmentation in surgery. Lightweight models achieve accurate near real-time performance, paving the way for clinical validation and integration of HSI-based intraoperative guidance into orthopedic workflows.</p> Graphical abstract <p></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

Hyperspectral imaging for orthopedic surgery: lightweight AI enables real-time bone and cartilage segmentation

  • 期刊论文

  • 发表日期：

    2026年3月25日

查看全文
  • Aya Hage Chehade,
  • Nadine Abdallah Saab,
  • Nesma Settouti,
  • Olga Assainova,
  • Chafiaa Hamitouche,
  • Mohammed El Amine Bechar,
  • Marwa El Bouz

摘要

Abstract

Accurate intraoperative identification of bone and cartilage is critical for orthopedic surgery, particularly during joint replacement procedures. Current approaches rely on direct visual inspection, which may be compromised by limited visibility or tissue deformation. This study represents the first investigation of hyperspectral imaging (HSI) for orthopedic surgery, assessing its potential for real-time intraoperative tissue segmentation when combined with lightweight artificial intelligence (AI) models, evaluated offline on in vivo porcine data. We leveraged the publicly available HeiPorSPECTRAL dataset, selecting Bone, Cartilage, and an (Others) class as representative orthopedic tissues. Pixel-based (multilayer perceptron (MLP), one-dimensional convolutional neural network (1D-CNN)) and patch-based (two-/three-dimensional U-Net (2D/3D U-Net)) models were systematically compared, with and without spectral normalization. Performance was assessed using accuracy, Dice similarity coefficient, and inference time. Normalization degraded performance across all models. The MLP achieved the best trade-off between accuracy (97.03%) and computational efficiency, segmenting a full hyperspectral cube in \(<1\) s. While 3D U-Net provided slightly higher Dice scores (0.96-0.99), inference required  8 s per cube, limiting clinical applicability. Qualitative inspection confirmed the robustness of the MLP in delineating bone and cartilage, even in challenging intraoperative scenarios. HSI combined with lightweight AI enables accurate and rapid tissue segmentation, highlighting its potential as a real-time intraoperative guidance tool in orthopedic surgery. Future work will focus on spectral channel reduction, in vivo validation, and integration with surgical instrumentation to accelerate clinical translation. We provide the first evidence that HSI can be applied to orthopedic tissue segmentation in surgery. Lightweight models achieve accurate near real-time performance, paving the way for clinical validation and integration of HSI-based intraoperative guidance into orthopedic workflows.

Graphical abstract