Photo Credit: Ankarb
The following is a summary of “Cross-site validation of lung cancer diagnosis by electronic nose with deep learning: a multicenter prospective study,” published in the May 2024 issue of Pulmonology by Lee et al.
Despite extensive investigation into using electronic nose (eNose) technology for diagnosing lung cancer, the challenge of cross-site validation has remained unaddressed, limiting its widespread adoption. To bridge this gap, researchers conducted a multi-center prospective study to validate eNose-based lung cancer diagnosis across two referral centers, a task never before undertaken in existing literature. Over the period from 2019 to 2022, patients with lung cancer, healthy controls, and diseased controls were recruited from both sites. Deep learning models were developed using a training cohort from one center and subsequently tested on a separate cohort from the other. To enhance model performance, the study group employed Semi-Supervised Domain-Generalized (Semi-DG) Augmentation (SDA) and Noise-Shift Augmentation (NSA) techniques, both with and without fine-tuning.
The study comprised 231 participants, with the training/validation cohort comprising 168 individuals and the test cohort comprising 63 individuals. While the model demonstrated satisfactory performance within the validation cohort from the same hospital, its direct application to the test cohort yielded suboptimal results. However, the implementation of data augmentation methods in the training cohort significantly improved performance, with SDA and NSA achieving impressive AUC values. Further enhancements were observed with the application of fine-tuning methods, resulting in notably higher AUC scores. The findings highlight the feasibility of achieving cross-site validation for eNose-based lung cancer diagnosis through the judicious use of data augmentation and fine-tuning techniques. Consequently, eNose breathprints present themselves as a promising, non-invasive, and potentially scalable solution for the detection of lung cancer.
Source: respiratory-research.biomedcentral.com/articles/10.1186/s12931-024-02840-z
