A reliable contour detection method for volatomics analysis with comprehensive two-dimensional gas chromatography leveraging image classification and instance segmentation.

Comprehensive Two-Dimensional Gas Chromatography (GC × GC) provides unparalleled resolving power for volatomics profiling. However, existing contour detection methods, when applied to these complex samples, frequently suffer from false negatives due to overlapped contours, severely hindering accurate quantification. Even when optimized, these methods struggle to resolve such ambiguities. This study proposes a hybrid deep learning framework to overcome false negatives caused by overlapped contours in GC × GC analysis by integrating image classification and instance segmentation. The approach involves four key steps: (1) constructing initial contour maps using an improved PeakCET v2 with a Laplacian operator; (2) classifying single and multi-peak contours with ResNet18; (3) segmenting overlapping contours via YOLO 11l; and (4) evaluating the segmentation results to identify singular contours. ResNet18 achieved a classification accuracy of 98.59%, outperforming other models with perfect precision. The YOLO 11l component demonstrated exceptional segmentation capability, attaining a mAP50 exceeding 87% and securing the highest mAP50-95 among tested architectures. Validation against diverse rose oil datasets confirmed the method's generalizability and robustness. By significantly reducing the dependency on manual curation, this ResNet18-YOLO 11l pipeline presents an automated, time-efficient solution for processing complex GC × GC data.