Performance-Enhanced Algorithm for Lung Cancer Screening Using Knowledge Distillation-Assisted Recurrent Neural Networks

Lung cancer is the leading cause of global cancer incidence and mortality. However, early diagnosis can significantly improve the postoperative survival rate. Featuring non-invasive, real-time, and affordable properties, breath analysis systems based on the electronic nose are highly promising compared to traditional diagnostic methods such as imaging and pathological examinations. Here we propose using a knowledge distillation-assisted recurrent neural network equipped with an attention mechanism (KD-ARNN) to improve the overall performance of the electronic nose. Specifically, to overcome the bottleneck of overfitting caused by the extremely limited number of lung cancer samples, a multi-database fusion learning strategy based on knowledge distillation was introduced. By transferring prior knowledge from the teacher model to the student model, cross-domain knowledge transfer was achieved from the gas concentration regression task to the disease state classification task, significantly enhancing the model’s generalization capability under limited sample conditions. On a self-built lung cancer breath gas dataset comprising 89 samples, the KD-ARNN model demonstrated superior performance compared to multiple machine learning and neural network methods. Even in an extremely small-sample scenario with only 30 samples, the model maintained robustness. Furthermore, preliminary interpretability analysis based on KD-ARNN revealed the relative contributions of different gas sensors in lung cancer identification, further substantiating the rationality of the model’s decision-making. The results indicate that KD-ARNN, as a reliable and efficient breath analysis tool, holds significant potential for early screening of lung cancer and other diseases.