CZII CryoET Object Identification Challenge - 5th place solution - Youssef Ouertani

Deposition Overview

The 5th place solution employs a compact DeepFinder-like 3D neural network with a 3D batch normalization layer at the input and reduced channel counts for particle detection in cryoET data. The model is trained using label smoothingcross-entropy loss with smoothing factor of 0.01, with binary masks created using customized particle radius factors of 0.5 or less. The inference pipeline splits volumes into patches with minimal overlap, applies 6-way test-time augmentation (3 flips and 3 rotations), and uses connected components analysis with area-based filtering for post-processing.