Optimal patch size for LArTPC pixel-level classification

Determine the optimal N×N patch size around each hit for pixel-level classification of track-like versus shower-like energy deposits in liquid argon time projection chamber (LArTPC) images, balancing contextual information against noise and irrelevant structures in wire-plane projections such as those from the MicroBooNE detector.

Background

The paper classifies each pixel (hit) in LArTPC images as track-like or shower-like using local patches centered on the pixel. Larger patches provide more context about the surrounding topology, which can aid classification, but may also introduce unrelated structures that degrade performance.

Using the MicroBooNE dataset, the authors evaluate how patch size impacts model performance across classical and quantum-enhanced architectures. While large models can exploit larger patches effectively, smaller models may be confused by excessive context. The authors explicitly note that identifying the optimal patch size is still unresolved.