Research on lung tumor boundary diagnosis method based on laser-induced breakdown spectroscopy combined with deep learning

To evaluate the potential of using a deep learning network model combined with Laser-Induced Breakdown Spectroscopy (LIBS) for the rapid diagnosis of lung tumor tissue boundaries, LIBS was employed to rapidly detect lung tumors and boundary tissues from 45 patients. Multivariate analysis was used to select 12 characteristic spectral lines from 6 elements as the input variables for the model. To address the challenges of complex preprocessing and feature extraction in biological tissue spectral data, a deep learning-based spectral feature processing framework using the ResNet18 backbone was designed. This framework was then evaluated against 3 machine learning models and 2 classical deep learning network architectures. The results indicate that the ResNet18 network model demonstrates significantly superior recognition capabilities compared to other models. It achieved accuracy, sensitivity and specificity of 99.6%, 100% and 99.3%, respectively, with precision and recall both reaching 99.6%. This suggests that, while maintaining high recognition accuracy, the model exhibits balanced recognition capabilities across various spectral data types, thereby minimising missed detections and maximising identification of all positive cases. Cross-dataset testing revealed that the ResNet18 model established in this study possesses robust cross-sample set generalisation capabilities, achieving 98.6% accuracy. This study indicates that the combination of LIBS spectroscopy and deep learning shows preliminary potential for distinguishing lung tumor from boundary tissue and may serve as a potential auxiliary screening technique, providing preliminary spectral evidence for the rapid screening of lung tumor margins.