Diagnosis model for mouse lung cancer based on terahertz spectroscopy and a transformer network

Terahertz time-domain spectroscopy technology offers a non-destructive and non-invasive approach to elucidating microstructure and chemical properties of cancer tissues, providing significant advantages for early diagnosis and monitoring. However, challenges such as the presence of outliers, sample class imbalance, and limited diagnostic accuracy persist. To enhance dataset quality, the boxplots were applied to identify and remove abnormal data, and the Savitzky-Golay filter was employed for spectral denoising in this paper. A high-quality dataset comprising 1,028 spectra from normal mouse lung tissues and 1,547 spectra from mouse lung cancer tissues was obtained. To address the shortage of normal lung tissue samples, a Wasserstein generative adversarial network with gradient penalty was utilized. Furthermore, to improve diagnostic accuracy, we developed a novel diagnostic model, MSFPT-Net. This model integrated seven distinct terahertz spectral features and employed a feature pyramid network to extract multi-scale representations, alongside a Transformer network to capture long-range dependencies within spectral sequences. The processed features were subsequently input into a fully connected network to achieve precise lung cancer diagnosis. Experimental results indicate that MSFPT-Net achieves outstanding performance across six evaluation metrics and opens new possibilities for lung cancer diagnosis and treatment.