Real-Time Detection of Trace Breath Isoprene Based on Circular Domain Spectral Reconstruction Filtering Combined with Convolutional Neural Network

The detection of trace isoprene in breath provides a noninvasive method for lung cancer diagnosis. However, the presence of interfering components and the parts per billion (ppb) concentration levels of isoprene in breath complicate detection. In this study, we propose an optical sensor based on circular domain reconstruction filtering and convolutional neural network (CNN), enabling the real-time detection of breath isoprene using ultraviolet differential optical absorption spectroscopy (UV-DOAS) for the first time. First, we obtained the differential absorption spectra of isoprene using UV-DOAS and analyzed the impact of interfering components including water vapor (H₂O) on the spectral characteristics. Second, we proposed a novel circular domain reconstruction filtering method that effectively mitigates noise and removes interference from components including ammonia (NH₃) and nitric oxide (NO) by discretizing disturbance absorption features. By mapping the absorption features to the circular domain, the proposed filtering method eliminates discrete noise and interference, providing a novel perspective on trace gas detection and spectral analysis. Based on the filtered spectra, a CNN model was constructed to invert isoprene concentration. Laboratory results show that the sensor has a detection limit of 3.98 ppb·m and provides accurate and real-time breath isoprene sensing ranging from 21.32 to 1254.20 ppb. Test results from human samples further demonstrate the effectiveness of the sensor in detecting trace isoprene in breath. Our sensor not only shows potential for application in isoprene detection but also advances the use of broadband spectroscopy in breath analysis.