Segmentation of gastroesophageal reflux events using a semi-U-Net architecture with 1D/2D CNNs

U-Net has gained traction in biomedical signal processing, particularly for segmenting 1D waveforms. Building on this success, we propose a U-Net-inspired architecture that integrates both 2D and 1D CNNs to effectively learn and segment gastroesophageal reflux (GER) events from Multichannel Intraluminal Impedance (MII) signals-specifically, a 6-channel 1D impedance signal. Current methods for GER detection are limited by the absence of efficient software, leading to time-consuming manual interpretation that is prone to errors. As a key contribution, we are also releasing the dataset of MII signals and GER annotations publicly to facilitate further research and algorithm development. In our architecture, the 2D CNN serves as the first encoder in a semi-U-Net structure to capture features across all channels. Subsequently, all other encoders and decoders utilize 1D CNNs to preserve the 1D nature of the signal while minimizing the number of parameters. After network training, the model segments GER areas in the 6th channel, utilizing a post-processing unit that accurately segments GER areas across all six channels. This unit ensures that selected GER events align with clinically defined criteria. The proposed architecture is compact and efficiently utilizes parameters, demonstrating strong generalizability across diverse GER events, with average durations of 17.52 ± 6.39 s. Outperforming existing methods, our approach achieves a sensitivity of 95.24% and a positive predictive value of 100%, indicating superior segmentation quality. We evaluated the model's robustness using data from 202 episodes containing 208 GER events collected from 26 patients who underwent 24-h MII pH monitoring. This semi-U-Net architecture, with its low parameter count, offers robust generalizability and adaptability to varying input durations. By improving GER event segmentation, our approach enhances the utility of 24-h MII-pH monitoring, enabling clinicians to make better-informed decisions for patient selection in invasive surgical procedures.