2.5D U-Net with Depth Reduction for 3D CryoET Object Identification

TL;DR

This work tackles automated protein molecule identification in cryoET tomograms, evaluated under the $F_\beta$ metric with $\beta=4$ to emphasize recall. It proposes a heatmap-based keypoint detection framework built on two 2.5D U-Net models with depth reduction, trained to output 3D heatmaps of particle centers, and employs a Bayesian-inspired ensemble with windowed inference and TensorRT acceleration. Ground-truth heatmaps are created by placing Gaussians at particle centers, with an offset of $1.0$ when converting coordinates and model-specific Gaussian widths. The method uses distinct loss functions to handle extreme class imbalance, performs multi-model inference with cross-model averaging, and applies NMS-based post-processing and coordinate remapping to final centers. The approach achieved 4th place in CZII competition, and ablation results highlight the importance of window overlap and location-aware weighting, suggesting a practical pathway for scalable automated cryoET analysis.

Abstract

Cryo-electron tomography (cryoET) is a crucial technique for unveiling the structure of protein complexes. Automatically analyzing tomograms captured by cryoET is an essential step toward understanding cellular structures. In this paper, we introduce the 4th place solution from the CZII - CryoET Object Identification competition, which was organized to advance the development of automated tomogram analysis techniques. Our solution adopted a heatmap-based keypoint detection approach, utilizing an ensemble of two different types of 2.5D U-Net models with depth reduction. Despite its highly unified and simple architecture, our method achieved 4th place, demonstrating its effectiveness.

Figures (3)

• Figure 1: The architecture of yu4u's model.
• Figure 2: The architecture of tattaka's model.
• Figure 3: Sliding overlapping windows used in inference.