CZII CryoET Object Identification Challenge - 10th place solution - Josef Slavicek

Deposition Overview

The 10th place solution employs an ensemble of nine 3D UNets, pretrained on simulated data and finetuned on competition data, using a combination of Tversky loss and multiclass cross-entropy as the training objective. Predictions are merged by averaging logits equally across models, followed by thresholding to obtain detection regions and a custom KL-divergence based postprocessing approach to separate multiple particles detected within a single region. The solution achieves distinction through its postprocessing method that identifies particle centers by optimally placing theoretical particle probability density functions to minimize KL divergence with predicted density functions, further refined by using circumcenters of randomly selected boundary points to improve centroid predictions.

Annotation Methods Summary