CZII CryoET Object Identification Challenge - 6th place solution - tomoon33

Deposition Overview

The 6th place solution employs a combination of 2.5D UNet and 3D UNet architectures with various backbones (including SegResNet and DynUnet) to predict segmentation maps for particle detection in cryoET data. FocalTversky++ loss is utilized during training on 64×128×128 patches, with models being pretrained on custom-simulated data generated using Polnet and accelerated during inference via TensorRT conversion. A key innovation involves designing prediction maps that reflect model confidence rather than binary outputs, enabling effective ensemble averaging across 10 models.