High-precision identification of breast cancer based on end-to-end parallel spectral convolutional neural network assisted laser-induced breakdown spectroscopy

Abstract

Breast cancer (BC) continues to be a significant cause of morbidity and mortality among women globally, underscoring the critical need for efficient and accurate screening methods. In this study, we introduce a Parallel Spectral Convolutional Neural Network (PSCNN), an end-to-end model, to simultaneously perform laser-induced breakdown spectroscopy (LIBS) spectral preprocessing and BC identification. PSCNN demonstrated superior performance compared to traditional single-task models. In the spectral preprocessing task, the signal-to-background ratio and signal-to-noise ratio of the preprocessed spectra improved by 8.6 and 1.6 times, respectively, compared to the raw spectra. For the classification task, the PSCNN achieved a classification accuracy of 90% on 52 test blood plasma samples, surpassing the 78% accuracy of the principal component analysis with linear discriminant analysis (PCA-LDA) model and the 82% accuracy of a single-task deep CNN. Furthermore, the PSCNN classification results were corrected according to the source of the donor individual, where the accuracy, specificity, and sensitivity achieved 92%, 96%, and 89%, respectively, for distinguishing between BC and healthy control (HC) donors. Ablation experiments revealed that removing the preprocessing module of the PSCNN led to decreased overall model performance and overfitting, indicating that information sharing occurred between the two modules. The spectral preprocessing module introduced regularization constraints for the classification module, enabling the model to learn more effective features. Overall, the PSCNN enhanced the discrimination performance in BC spectral analysis through multi-task modeling.