CZII CryoET Object Identification Challenge - 8th place solution - I Cryo Everyteim

Deposition Overview

The 8th place solution employs an ensemble of four 3D U-Net model soups trained with different model sizes, parameters, and training data types, featuring patch sizes of (128,128,128) for training and (160,384,384) for inference with 25% overlap using Gaussian reconstruction to handle border artifacts. The models are pretrained on six synthetic tomograms with Gaussian denoising and fine-tuned using Dice cross entropy loss with AdamW optimization, incorporating geometric data augmentations including flipping and transposition during both training and test-time. The key innovation lies in the model soup approach, which averages weights from multiple models trained from the same pretrained weights, combined with their inference strategy using maximally large patchsizes and watershed segmentation for post-processing.

Annotation Methods Summary