The Utility of Deep Learning–Based Segmentation in Automated Scintillator Tile Dimensioning

The Electron-Ion Collider (EIC), set to begin data collection in the 2030s, is to be built at Brookhaven National Laboratory. Through high-precision electron-ion collisions, this "super microscope" will probe the internal structure of protons and nuclei. Among the multiple sub-detectors that enable this process, the Longitudinally-Segmented Forward Hadronic Calorimeter (LFHCal), located in the forward rapidity region, plays a key role. Thousands of small scintillator tiles will be used to provide high spatial resolution in this region, enabling accurate measurement and reconstruction of the energy of particles produced in collisions. To ensure consistent data quality and detector performance, the tile dimensions must conform to strict design specifications. As an alternative to the time-consuming and error-prone process of manual validation, this project investigates deep learning and image segmentation techniques - specifically U-Net and Mask-RCNN - to automate the task. By evaluating segmentation performance across Average Precision, Boundary IoU, and Dice Coefficient metrics, we fine-tune the training pipeline, hyper-parameters, and architecture to improve model accuracy and efficiency. This poster presents the results of that evaluation and discusses how these methods can be extended to other detector components.