Application of an interpretable Raman-Spectral classification model based on a transposed convolution and attention mechanism to tumor-tissue screening

Abstract

Skin and gastric cancers are known for their high mortality rates. In contemporary clinical practice, medical imaging methods are prevalently utilized for the analysis and evaluation of tumor tissues. However, these methods are limited by two major factors. Firstly, they heavily rely on the experience of personnel, which may lead to diagnostic inaccuracies. Secondly, the significant changes in cell morphology that these methods detect typically manifest in the middle or later stages of the disease, rendering them less effective for early cancer screening. Interestingly, these diseases are often marked by alterations in metabolic products, which can be used to identify early-stage cancerous tissues based on compositional changes. This approach not only significantly reduces reliance on experiential judgment but also minimizes manual intervention. Employing Raman spectroscopy, an optical testing method known for its molecular fingerprinting capabilities, this paper introduces the TA-Net classification model. Developed based on Raman spectroscopy data, TA-Net outperforms conventional models in terms of accuracy, specificity, and interpretability. It incorporates a transposed convolution and attention mechanism to address the challenge of varying data lengths collected by different Raman spectrometers, a common issue that has hindered the application of previously trained models. The model features down-sampling, up-sampling, convolutional attention mechanisms, and global-pooling classification modules, and is trained end-to-end. This reduces preprocessing complexity and results in high-precision, interpretable prediction outputs. In comparative experiments using collected skin and gastric cancer Raman spectral data, the TA-Net model demonstrated improvements in sensitivity (0.74% and 6.12%), specificity (5.63% and 5.09%), and accuracy (3.15% and 5.54%) over the suboptimal model. This study offers a promising approach for early cancer screening and detection, holding significant practical value for clinical diagnosis.