Yousuf Soha, Karukappadath Mohamed Irfan, Zam Azhar
Division of Engineering, Laboratory for Advanced Bio-Photonics and Imaging (LAB-π), New York University Abu Dhabi, Abu Dhabi, United Arab Emirates.
Department of Biomedical Engineering, Department of Electrical and Computer Engineering, Tandon School of Engineering, New York University, Brooklyn, New York, USA.
Lasers Surg Med. 2025 Aug;57(6):517-527. doi: 10.1002/lsm.70031. Epub 2025 May 28.
Integrating machine learning with Raman spectroscopy (RS) shows strong potential for intraoperative guidance in orthopedic procedures, but limited algorithm transparency remains a barrier to clinician trust. This study aims to develop interpretable machine learning models capable of accurately classifying bovine tissue types (bone, bone marrow, fat, and muscle) relevant to orthopedic surgery by identifying key Raman biomarkers to improve model transparency.
A portable RS system equipped with a 785 nm fiber-optic probe was used to collect spectral data from excised bovine tissues, including bone, bone marrow, muscle, and fat. One-dimensional convolutional neural network (1D-CNN) and support vector machine (SVM) models were developed to classify these tissue types. The Raman spectral data were divided using a sample-based, stratified splitting strategy and evaluated across 30 independent iterations. Feature importance maps were generated for both models, and matching scores were calculated to correlate significant spectral features with known Raman biomarkers.
Through feature importance analysis and matching scores generated by the 1D-CNN and SVM models, critical Raman biomarkers-including hydroxyapatite, lipids, amino acids, and collagen-were identified as essential for distinguishing between the different bovine tissue types, providing deeper insights into their molecular differences.
The integration of interpretable machine learning models with RS enabled accurate differentiation of bovine tissues relevant to orthopedic surgery, while enhancing model transparency through biomarker identification. Linking model predictions to biologically meaningful Raman features supports the development of RS as a reliable tool for precision-guided surgical procedures.
将机器学习与拉曼光谱（RS）相结合在骨科手术的术中指导方面显示出强大潜力，但算法透明度有限仍是临床医生信任的障碍。本研究旨在开发可解释的机器学习模型，通过识别关键拉曼生物标志物来准确分类与骨科手术相关的牛组织类型（骨、骨髓、脂肪和肌肉），以提高模型透明度。
使用配备785 nm光纤探头的便携式RS系统从切除的牛组织（包括骨、骨髓、肌肉和脂肪）中收集光谱数据。开发了一维卷积神经网络（1D-CNN）和支持向量机（SVM）模型来对这些组织类型进行分类。拉曼光谱数据采用基于样本的分层拆分策略进行划分，并在30次独立迭代中进行评估。为两个模型生成特征重要性图，并计算匹配分数以将重要光谱特征与已知拉曼生物标志物相关联。
通过1D-CNN和SVM模型生成的特征重要性分析和匹配分数，确定了关键拉曼生物标志物，包括羟基磷灰石、脂质、氨基酸和胶原蛋白，它们对于区分不同的牛组织类型至关重要，为深入了解它们的分子差异提供了依据。
可解释的机器学习模型与RS的整合能够准确区分与骨科手术相关的牛组织类型，同时通过生物标志物识别提高模型透明度。将模型预测与具有生物学意义的拉曼特征联系起来，支持将RS开发成为精确引导手术的可靠工具。
